Catalog / Business processes / Revenue operations / Process-to-Adoption / Instrument a Process
reference process · revenue-operations · 16 activities · 12 on the rosterInstrument a Process
A process is written down, and the systems are changed so that departing from it shows up on its own. The run picks the steps worth watching, says what a followed step and a skipped step look like in the data, finds the evidence that already exists, and builds the fields and events for the evidence that does not. A privacy check runs before anything is watched, and the people being measured are told by a person rather than by a report. The thresholds are set, the detection is tested against runs whose answer is already known, and the numbers go to the review.
The walk
The document
One run
Adoption
The activities
What happens in a run
16 activities from a process publishes at a version, or its measure is asked for to departure from the process is visible without anybody being asked, 4 of them gates a person has to sign. Drag the diagram to move along it.
Instrument a Process a process publishes at a version, or its measure is asked for → departure from the process is visible without anybody being asked
departure cannot be seen in any system the watching is not allowed the change is refused or cut back the detection misses known departures the signal fires on runs that followed the step was never observable a process publishes at a version, or its measure is asked for 1 Take in the Published Pro… the process document, at the version being measured 2 Pick the Steps Worth Watc… which steps matter enough to watch, and which do not 3 Say What Departure Looks … what a followed step and a skipped step look like a person signs · never an agent 4 Find Where the Evidence A… the fields and logs that already show the behavior 5 Name the Signals That Are… the departures no system would currently show 6 Design the System Changes the fields, events and logs that have to change 7 Check the Watching Agains… what may be watched here, and what may not 8 Get the Change Approved the systems owner signs for the change a person signs · never an agent 9 Build the Fields and the … the fields and events are made in the systems 10 Build the Baseline what following the process looked like before 11 Test the Detection on Kno… known good and known bad runs put through detection 12 Set the Thresholds how much departure is worth raising, and to whom a person signs · never an agent 13 Turn On the Report and th… the report and the alert start arriving 14 Tell the People They Are … the people doing the work hear it from a person a person signs · never an agent 15 Watch the First Weeks false alarms are found and the detection is tuned 16 Hand the Measure to the R… the numbers become the input to the adoption review departure from the process is visible without anybody being asked The roster this process needs
Hover a name to see the activities it holds. A dashed one is a person, and stays one.
The document
The document, with the blanks marked
Everything in amber is yours to fill in: who owns it, when it takes effect, which platforms, which numbers, and who holds each activity. Everything else is the process, and it is the same wherever it is run.
PROCESS: instrument a process id: <team> /instrument-a-process v1
from: ref/rev/instrument-a-process v1
owner: <who> effective: <date>
trigger: a process publishes or amends at a version in <your process
register>, or <who> asks for an existing process to be measured
watch: record=<process document>
system=<your process register>
change=<a process publishes or amends at a version>
or watch: record=<the request to measure>
system=<where requests are raised>
change=<somebody asks for an existing process to be
measured>
concurrency: runs may overlap - <how many> processes are being
instrumented at once, and one process is instrumented by
one run at a time
goal: every step the owner named as worth watching produces evidence on
its own, the evidence separates a followed run from a departed
one, and the numbers reach the review on a fixed cadence
phases:
take-in-process - human: read the published process at the version
being measured, and the steps it says are gates
owner: process-librarian
after: trigger
automation: assisted
pick-steps - convenes decide-and-announce: which steps matter
enough to watch. The owner names them, and the
steps left unwatched are written down as
deliberately unwatched
owner: adoption-monitor
after: take-in-process
by: <days>
automation: <level>
define-departure - human: for each watched step, what a followed
step looks like in the data and what a skipped
one looks like, in words a person can check
owner: adoption-monitor
after: pick-steps
automation: never
find-evidence - runs collect-and-report: the fields, timestamps
and logs that already carry the behavior, each
named with the system it sits in
owner: analytics
after: define-departure
by: <days>
automation: supervised
name-gaps - runs assessment: the departures that no system
would show today, and what it would take to see
each one
owner: adoption-monitor
after: find-evidence
automation: <level>
design-changes - human: the fields, events and log lines that
have to change, written as a change request
against <your systems>
owner: adoption-monitor
after: name-gaps
by: <days>
automation: assisted
check-policy - runs assessment: the proposed watching scored
against <your privacy policy> and <your
employment agreements>, per field
owner: standards-keeper
after: design-changes
by: <days>
automation: <level>
approve-change - convenes approval: the systems owner signs the
change, and the privacy reviewer signs the
watching, as named signers
owner: <your systems owner role>
after: check-policy
by: <days>
automation: never
build-instruments - human: the fields, the events and the logging
are made in the systems, at a version, with a
way to turn each one off
owner: <your systems administrator role>
after: approve-change
by: <days>
automation: <level>
build-baseline - runs collect-and-report: what following the
process looked like over the last <how many>
periods, from the evidence that already existed
owner: analytics
after: find-evidence
by: <days>
automation: supervised
test-detection - runs assessment: <how many> runs whose answer is
already known, good and bad, put through the
detection and scored against that answer
owner: adoption-monitor
after: build-instruments + build-baseline
automation: <level>
set-thresholds - convenes decide-and-announce: how much departure
is worth raising, to whom, and how fast. <who>
confirms the numbers
owner: decision-coordinator
after: test-detection
automation: never
turn-on-signal - human: the report goes out on <your cadence> and
the alert path opens, both naming the process
and the version they measure
owner: reporting-manager
after: set-thresholds
automation: <level>
tell-the-measured - convenes briefing: the people doing the work
hear from their own manager what is measured,
what is not, and who sees the numbers
owner: <your process owner role>
after: set-thresholds
by: <days>
automation: never
watch-first-weeks - runs collect-and-report: every signal raised in
the first <how many> weeks, checked by hand
against the run it fired on
owner: adoption-monitor
after: turn-on-signal + tell-the-measured
by: <weeks>
automation: <level>
hand-to-review - human: the measure, its baseline and its known
blind spots handed to the adoption review as its
standing input
owner: reporting-manager
after: watch-first-weeks
automation: <level>
run-scoped:
false-alarms - runs collect-and-report owner: adoption-monitor
every: <cadence>
from: turn-on-signal until: run close
status - runs roll-call owner: adoption-monitor
every: <cadence>
from: pick-steps until: turn-on-signal
handoffs:
take-in-process -> pick-steps [published-process]: the process at its
version, with every step it calls a gate marked as a gate
pick-steps -> define-departure [watched-steps]: the watched steps, and
the unwatched ones with the reason each was left out
define-departure -> find-evidence [departure-signs]: for each watched
step, the observable difference between a followed run and a
departed one
find-evidence -> name-gaps [existing-evidence]: the fields and logs
that already carry the behavior, each naming its system and how far
back it goes
find-evidence -> build-baseline [baseline-fields]: the same fields, so
the baseline is built from the evidence that existed before the
change
name-gaps -> design-changes [evidence-gaps]: the departures nothing
would show, and what each one would cost to see
design-changes -> check-policy [change-request]: the change request,
field by field, saying what each new field records about a person
check-policy -> approve-change [policy-score]: the scored request,
with the policy line behind every refusal
approve-change -> build-instruments [approved-change]: the approved
change at the signed version, and the fields the privacy reviewer
struck out
build-instruments -> test-detection [live-instruments]: the fields and
events live, with a dated test showing each one arriving
build-baseline -> test-detection [baseline]: what compliance looked
like before, so a change in the number is not read as a change in
behavior
test-detection -> set-thresholds [detection-score]: the score against
the known runs, with every miss and every false alarm listed
set-thresholds -> turn-on-signal [thresholds]: the numbers, who is
raised to, and how fast
set-thresholds -> tell-the-measured [what-is-measured]: the same
numbers, in the words the people doing the work will hear them in
turn-on-signal -> watch-first-weeks [live-signal]: the report and the
alert path live, each naming the process version it measures
watch-first-weeks -> hand-to-review [tuned-measure]: the tuned
detection, its false alarm rate, and the steps it still cannot see
deviations:
name-gaps -> define-departure [no-system-shows-departure]: departure
cannot be seen in any system, so what departure looks like is
written again or the step moves to the unwatched list
check-policy -> pick-steps [watching-refused]: the policy check
refuses the watching, so the steps worth watching are picked again
from what is left
approve-change -> design-changes [change-cut-back]: the systems owner
refuses or cuts back the change, so the fields and events are
designed again
test-detection -> design-changes [detection-misses-departures]: the
detection misses departures it was shown, so it is designed again
and the same known runs are put through it
watch-first-weeks -> set-thresholds [signal-fires-wrongly]: the signal
fires on runs that followed the process, so the thresholds are set
again and the runs that fired wrongly are named
watch-first-weeks -> define-departure [step-never-observable]: the
first weeks show a watched step nothing can see, so what departure
looks like for that step is written again
watch-first-weeks -> build-instruments [instrument-broke]: a field is
renamed or a screen retired and the evidence stops arriving, so the
instruments are built again rather than the drop being read as
perfect compliance
bindings:
roster: <who holds each role - the agent claiming adoption-monitor,
the agent claiming analytics, the named systems owner, the
privacy reviewer, and the managers who tell their teams>
systems: the CRM (write), <your systems> (write),
<your logging system> (read), <your reporting tool> (write),
the alert path (trigger), the process register (read)
data: the process document <version> , <your privacy policy>
<version> , <your employment agreements> , the baseline period
policy:
nothing is watched that <your privacy policy> does not allow, and the
check runs before the field is built, not after
the people doing the work are told what is measured before the first
report goes out, and they are told by their own manager
a measure names the process version it measures. A process that amends
sends its measure back through test-detection
a signal that has not been checked against a real run in <how many>
weeks is switched off rather than left firing
the report shows a rate and a count, never a ranked list of named
people, unless <who> asks for names in writing
every instrument can be turned off without a deployment
measures:
cycle time: <target> from the published process to the first report
coverage: <target> share of the watched steps that produce evidence
without anybody being asked
precision: <target> share of raised signals that were real departures
quality gate: no field is built before the policy check passes, and no
report goes out before the people in it were told
Copy
Take it somewhere
Use this process in LangGraph
close
Paste this into an assistant that can read the web, such as Claude, ChatGPT or Cursor. It reads the specification and the current LangGraph documentation, then writes two files: the graph, and a note on what did not survive the translation. Read the note first. What a runtime cannot express is the part worth arguing about, and this process is a draft to argue with.
copy the prompt
343 lines · the document is inside it, so nothing else is needed
Convert the business process below into a runnable LangGraph graph:
one Python file with a TypedDict state, a StateGraph, nodes, edges,
conditional edges and a checkpointer.
The document is a reference process written to the Agent Processes
specification. Read the specification before you start, because it defines
terms that look ordinary and are not:
https://agentcatalog.com/spec/agent-processes
Sections 6 (the phase graph), 6.5.1 (exception edges), 6.7.1 (deviations),
7 (automation) and 8 (handoffs) are the ones this conversion turns on.
Then read the current documentation for the primitives you will need, rather
than relying on what you remember of the API:
https://docs.langchain.com/oss/python/langgraph/interrupts
interrupt() and Command(resume=), which is how a gate stops a run
https://reference.langchain.com/python/langgraph/graph/state/StateGraph
StateGraph, add_edge, add_conditional_edges, defer
WHAT THE DOCUMENT ASKS FOR
These hold wherever the process lands, and they matter more than style.
1. Each phase under `phases:` becomes one step, and keeps its name.
2. `after:` gives the edges. `after: a + b` is a join and waits for BOTH.
Reading it as "either" is the defect the specification calls out by name.
3. Every handoff carries a key in square brackets. Each key becomes one field
on the run's state, named exactly as the key with hyphens turned into
underscores, and the sentence beside it becomes that field's comment. The key
is the stable name; the sentence is prose that may be rewritten.
4. A phase MUST NOT begin before its inbound handoff exists. Where that is
checkable, check it in the step rather than assuming it.
5. `automation: never` is a gate a person signs. The run stops there and does
not continue until a person's decision comes back. Do not turn one into a
notification, a log line, or an automatic transition, whatever the queue
looks like.
6. Each line under `deviations:` is a backward or sideways edge, returning to
the phase named on the right. The key in brackets names it, and that name
belongs in the code.
7. A phase whose `after:` reads like "X or Y, whichever could not finish" is an
exception edge: it is entered when those phases FAIL, not when they succeed.
Do not wire it as an ordinary successor.
8. Anything in angle brackets is a blank the adopting organization fills in.
Leave each one as a named constant at the top of the file with a TODO. Do not
invent a value, a threshold or a date.
9. Record the document's `from:` line at the top of the file, so it says which
reference process and which version it was generated from.
10. Run-scoped lines under `run-scoped:` are work that runs alongside the whole
process rather than at one point in it, and a run may not close while one is
unfinished. Say in the code what you did about them, including if the answer
is that the runtime has nowhere to put them.
HOW THAT LOOKS IN LANGGRAPH
11. A phase is a node added with `add_node`, under the phase's own name.
12. The state is a TypedDict. Each handoff key is one field on it.
13. A join is the trap. `add_edge(["a", "b"], "c")` looks right and releases
once: when a backward edge re-enters ONE arm, the joined node never runs
again, and the run ends early reporting success rather than raising. Mark the
joined node `defer=True` and re-check inside it that both inbound handoffs
exist.
14. A gate is `interrupt()` inside the node, resumed with `Command(resume=...)`.
The platform lets anything at all call resume, so require the resumed value to
name a person and a date and refuse anything else. Say in the fidelity note
that this proves only that whoever resumed typed a name, because
`Command(resume=True)` from a scheduled job is indistinguishable from a person
signing.
15. A deviation is `add_conditional_edges` with a routing function named after
the key in brackets.
16. Pass a durable checkpointer rather than taking the in-memory default. The
gates wait days, and the default loses every paused run on restart.
17. Leave every phase body unimplemented, raising until somebody registers an
implementation. The automation level is a blank, so writing a body would
answer on the adopter's behalf whether an agent may do that work.
Produce a second file alongside it, `FIDELITY.md`, and treat it as the more
important of the two. The code is for whoever builds this. The fidelity note
is for whoever has to decide whether this platform suits the process at all,
and that is usually a different person who will never read the code.
It has three parts.
**What came across.** Briefly: how many phases became steps, how many handoff
keys became state fields, which gates stop the run, which deviations became
edges. Counts and names, not reassurance.
**What did not, and what was done instead.** One entry per gap. For each one,
say what the document requires, what the platform can actually express, what
you did in its place, and what breaks if somebody later removes your
workaround. This last part matters most: a workaround nobody understands is a
workaround somebody deletes.
**What a person still has to decide.** The blanks are not a translation
failure, they are the point of a reference process, so list what has to be
filled in before this could run against anything real, and say which of those
choices the platform constrains.
Write it in plain English for somebody who has not read the specification, and
do not soften it. A translation of a reference process is a draft to argue
with, not a build artifact, and the honest account of what was lost is the most
useful thing you will produce.
Here is the process document.
```
PROCESS: instrument a process id: <team>/instrument-a-process v1
from: ref/rev/instrument-a-process v1
owner: <who> effective: <date>
trigger: a process publishes or amends at a version in <your process
register>, or <who> asks for an existing process to be measured
watch: record=<process document>
system=<your process register>
change=<a process publishes or amends at a version>
or watch: record=<the request to measure>
system=<where requests are raised>
change=<somebody asks for an existing process to be
measured>
concurrency: runs may overlap - <how many> processes are being
instrumented at once, and one process is instrumented by
one run at a time
goal: every step the owner named as worth watching produces evidence on
its own, the evidence separates a followed run from a departed
one, and the numbers reach the review on a fixed cadence
phases:
take-in-process - human: read the published process at the version
being measured, and the steps it says are gates
owner: process-librarian
after: trigger
automation: assisted
pick-steps - convenes decide-and-announce: which steps matter
enough to watch. The owner names them, and the
steps left unwatched are written down as
deliberately unwatched
owner: adoption-monitor
after: take-in-process
by: <days>
automation: <level>
define-departure - human: for each watched step, what a followed
step looks like in the data and what a skipped
one looks like, in words a person can check
owner: adoption-monitor
after: pick-steps
automation: never
find-evidence - runs collect-and-report: the fields, timestamps
and logs that already carry the behavior, each
named with the system it sits in
owner: analytics
after: define-departure
by: <days>
automation: supervised
name-gaps - runs assessment: the departures that no system
would show today, and what it would take to see
each one
owner: adoption-monitor
after: find-evidence
automation: <level>
design-changes - human: the fields, events and log lines that
have to change, written as a change request
against <your systems>
owner: adoption-monitor
after: name-gaps
by: <days>
automation: assisted
check-policy - runs assessment: the proposed watching scored
against <your privacy policy> and <your
employment agreements>, per field
owner: standards-keeper
after: design-changes
by: <days>
automation: <level>
approve-change - convenes approval: the systems owner signs the
change, and the privacy reviewer signs the
watching, as named signers
owner: <your systems owner role>
after: check-policy
by: <days>
automation: never
build-instruments - human: the fields, the events and the logging
are made in the systems, at a version, with a
way to turn each one off
owner: <your systems administrator role>
after: approve-change
by: <days>
automation: <level>
build-baseline - runs collect-and-report: what following the
process looked like over the last <how many>
periods, from the evidence that already existed
owner: analytics
after: find-evidence
by: <days>
automation: supervised
test-detection - runs assessment: <how many> runs whose answer is
already known, good and bad, put through the
detection and scored against that answer
owner: adoption-monitor
after: build-instruments + build-baseline
automation: <level>
set-thresholds - convenes decide-and-announce: how much departure
is worth raising, to whom, and how fast. <who>
confirms the numbers
owner: decision-coordinator
after: test-detection
automation: never
turn-on-signal - human: the report goes out on <your cadence> and
the alert path opens, both naming the process
and the version they measure
owner: reporting-manager
after: set-thresholds
automation: <level>
tell-the-measured - convenes briefing: the people doing the work
hear from their own manager what is measured,
what is not, and who sees the numbers
owner: <your process owner role>
after: set-thresholds
by: <days>
automation: never
watch-first-weeks - runs collect-and-report: every signal raised in
the first <how many> weeks, checked by hand
against the run it fired on
owner: adoption-monitor
after: turn-on-signal + tell-the-measured
by: <weeks>
automation: <level>
hand-to-review - human: the measure, its baseline and its known
blind spots handed to the adoption review as its
standing input
owner: reporting-manager
after: watch-first-weeks
automation: <level>
run-scoped:
false-alarms - runs collect-and-report owner: adoption-monitor
every: <cadence>
from: turn-on-signal until: run close
status - runs roll-call owner: adoption-monitor
every: <cadence>
from: pick-steps until: turn-on-signal
handoffs:
take-in-process -> pick-steps [published-process]: the process at its
version, with every step it calls a gate marked as a gate
pick-steps -> define-departure [watched-steps]: the watched steps, and
the unwatched ones with the reason each was left out
define-departure -> find-evidence [departure-signs]: for each watched
step, the observable difference between a followed run and a
departed one
find-evidence -> name-gaps [existing-evidence]: the fields and logs
that already carry the behavior, each naming its system and how far
back it goes
find-evidence -> build-baseline [baseline-fields]: the same fields, so
the baseline is built from the evidence that existed before the
change
name-gaps -> design-changes [evidence-gaps]: the departures nothing
would show, and what each one would cost to see
design-changes -> check-policy [change-request]: the change request,
field by field, saying what each new field records about a person
check-policy -> approve-change [policy-score]: the scored request,
with the policy line behind every refusal
approve-change -> build-instruments [approved-change]: the approved
change at the signed version, and the fields the privacy reviewer
struck out
build-instruments -> test-detection [live-instruments]: the fields and
events live, with a dated test showing each one arriving
build-baseline -> test-detection [baseline]: what compliance looked
like before, so a change in the number is not read as a change in
behavior
test-detection -> set-thresholds [detection-score]: the score against
the known runs, with every miss and every false alarm listed
set-thresholds -> turn-on-signal [thresholds]: the numbers, who is
raised to, and how fast
set-thresholds -> tell-the-measured [what-is-measured]: the same
numbers, in the words the people doing the work will hear them in
turn-on-signal -> watch-first-weeks [live-signal]: the report and the
alert path live, each naming the process version it measures
watch-first-weeks -> hand-to-review [tuned-measure]: the tuned
detection, its false alarm rate, and the steps it still cannot see
deviations:
name-gaps -> define-departure [no-system-shows-departure]: departure
cannot be seen in any system, so what departure looks like is
written again or the step moves to the unwatched list
check-policy -> pick-steps [watching-refused]: the policy check
refuses the watching, so the steps worth watching are picked again
from what is left
approve-change -> design-changes [change-cut-back]: the systems owner
refuses or cuts back the change, so the fields and events are
designed again
test-detection -> design-changes [detection-misses-departures]: the
detection misses departures it was shown, so it is designed again
and the same known runs are put through it
watch-first-weeks -> set-thresholds [signal-fires-wrongly]: the signal
fires on runs that followed the process, so the thresholds are set
again and the runs that fired wrongly are named
watch-first-weeks -> define-departure [step-never-observable]: the
first weeks show a watched step nothing can see, so what departure
looks like for that step is written again
watch-first-weeks -> build-instruments [instrument-broke]: a field is
renamed or a screen retired and the evidence stops arriving, so the
instruments are built again rather than the drop being read as
perfect compliance
bindings:
roster: <who holds each role - the agent claiming adoption-monitor,
the agent claiming analytics, the named systems owner, the
privacy reviewer, and the managers who tell their teams>
systems: the CRM (write), <your systems> (write),
<your logging system> (read), <your reporting tool> (write),
the alert path (trigger), the process register (read)
data: the process document <version>, <your privacy policy>
<version>, <your employment agreements>, the baseline period
policy:
nothing is watched that <your privacy policy> does not allow, and the
check runs before the field is built, not after
the people doing the work are told what is measured before the first
report goes out, and they are told by their own manager
a measure names the process version it measures. A process that amends
sends its measure back through test-detection
a signal that has not been checked against a real run in <how many>
weeks is switched off rather than left firing
the report shows a rate and a count, never a ranked list of named
people, unless <who> asks for names in writing
every instrument can be turned off without a deployment
measures:
cycle time: <target> from the published process to the first report
coverage: <target> share of the watched steps that produce evidence
without anybody being asked
precision: <target> share of raised signals that were real departures
quality gate: no field is built before the policy check passes, and no
report goes out before the people in it were told
```
Take it somewhere
Use this process in Microsoft Agent Framework
close
Paste this into an assistant that can read the web, such as Claude, ChatGPT or Cursor. It reads the specification and the current Agent Framework documentation, then writes two files: the workflow, and a note on what did not survive the translation. Read the note first. What a runtime cannot express is the part worth arguing about, and this process is a draft to argue with.
copy the prompt
398 lines · the document is inside it, so nothing else is needed
Convert the business process below into a runnable Microsoft Agent Framework workflow:
one Python file with Executor classes, a WorkflowBuilder, typed edges,
request_info gates and durable checkpoint storage.
The document is a reference process written to the Agent Processes
specification. Read the specification before you start, because it defines
terms that look ordinary and are not:
https://agentcatalog.com/spec/agent-processes
Sections 6 (the phase graph), 6.5.1 (exception edges), 6.7.1 (deviations),
7 (automation) and 8 (handoffs) are the ones this conversion turns on.
Then read the current documentation for the primitives you will need, rather
than relying on what you remember of the API:
https://learn.microsoft.com/en-us/agent-framework/workflows/human-in-the-loop
ctx.request_info, @response_handler, and answering a parked run later
https://learn.microsoft.com/en-us/agent-framework/workflows/checkpoints
what a checkpoint holds, and allowed_checkpoint_types
https://learn.microsoft.com/en-us/agent-framework/concepts/workflows/edges
add_edge with condition, add_fan_in_edges, add_switch_case_edge_group
https://learn.microsoft.com/en-us/agent-framework/concepts/workflows/state
ctx.set_state and ctx.get_state as they actually are today
WHAT THE DOCUMENT ASKS FOR
These hold wherever the process lands, and they matter more than style.
1. Each phase under `phases:` becomes one step, and keeps its name.
2. `after:` gives the edges. `after: a + b` is a join and waits for BOTH.
Reading it as "either" is the defect the specification calls out by name.
3. Every handoff carries a key in square brackets. Each key becomes one field
on the run's state, named exactly as the key with hyphens turned into
underscores, and the sentence beside it becomes that field's comment. The key
is the stable name; the sentence is prose that may be rewritten.
4. A phase MUST NOT begin before its inbound handoff exists. Where that is
checkable, check it in the step rather than assuming it.
5. `automation: never` is a gate a person signs. The run stops there and does
not continue until a person's decision comes back. Do not turn one into a
notification, a log line, or an automatic transition, whatever the queue
looks like.
6. Each line under `deviations:` is a backward or sideways edge, returning to
the phase named on the right. The key in brackets names it, and that name
belongs in the code.
7. A phase whose `after:` reads like "X or Y, whichever could not finish" is an
exception edge: it is entered when those phases FAIL, not when they succeed.
Do not wire it as an ordinary successor.
8. Anything in angle brackets is a blank the adopting organization fills in.
Leave each one as a named constant at the top of the file with a TODO. Do not
invent a value, a threshold or a date.
9. Record the document's `from:` line at the top of the file, so it says which
reference process and which version it was generated from.
10. Run-scoped lines under `run-scoped:` are work that runs alongside the whole
process rather than at one point in it, and a run may not close while one is
unfinished. Say in the code what you did about them, including if the answer
is that the runtime has nowhere to put them.
HOW THAT LOOKS IN MICROSOFT AGENT FRAMEWORK
11. Write for Python, and read those pages before writing a line. This API has
moved: `set_shared_state`, `RequestInfoExecutor`, `RequestInfoMessage` and
`send_responses_streaming` are all in the training data and none of them exist
any more. The .NET workflow API differs in kind rather than in spelling, so a
file written for one does not port by renaming.
12. A phase is a class deriving from `Executor` whose `super().__init__(id=...)`
takes the phase name verbatim. The id is not cosmetic: a checkpoint stores a
signature over the topology and the executor ids, so an id built from a run, a
timestamp or a counter cannot be resumed into.
13. Give each phase a `@handler` method and move work on with
`await ctx.send_message(...)`. A handler that returns without sending is a dead
end: the branch stops, the run converges, and it reports success. So a phase
you are deliberately leaving unimplemented must still send a placeholder
onward. Do not stub with `raise NotImplementedError`, which fails the run
instead of leaving it runnable.
14. `after: a` is `builder.add_edge(a, b)`. `after: a + b` is
`builder.add_fan_in_edges([a, b], target)`, and the target's handler must be
annotated `list[T]`, because a fan-in delivers one aggregated list rather than
the separate messages. A handler typed for the single value is dropped as a
mismatch with nothing raised.
15. The join is the trap here, and it is the opposite of LangGraph's. The
barrier re-arms: it clears its buffer when it fires and then demands a fresh
message from every source. So a deviation that re-enters ONE arm parks the
run forever waiting for an arm that will not run again, and the workflow ends
IDLE reporting success. Wherever a deviation re-enters one arm of a join,
replace the barrier with an ordinary edge from each arm into a small executor
that records each arrival with `ctx.set_state` and only forwards when every
expected key is present, and say in a comment that putting `add_fan_in_edges`
back reintroduces the stall.
16. A phase that is both a join target and a deviation target needs two
handlers, one annotated `list[T]` for the barrier and one annotated `T` for the
backward message. Write only the list handler and every backward edge into it
is discarded as a type mismatch, silently.
17. `automation: never` is `await ctx.request_info(request_data=...,
response_type=...)` inside the phase, answered by a `@response_handler` on the
same executor whose annotations match those exact types. The run parks at
`IDLE_WITH_PENDING_REQUESTS` and the host answers with
`workflow.run(stream=True, responses={request_id: value})`. If no handler
matches the pair, the framework logs a warning and parks anyway, so the gate
reads as working right up until somebody asks why the approval did not take.
18. A gate is only a gate if the wait survives a restart, so pass
`checkpoint_storage=FileCheckpointStorage(...)` to the builder. Checkpointing
is off by default and `InMemoryCheckpointStorage` reads as configured while
persisting nothing. Register every handoff payload type in
`allowed_checkpoint_types`, or the first restore raises. Anything an executor
keeps as an instance attribute is absent after a restore unless you export it
from `on_checkpoint_save` and read it back in `on_checkpoint_restore`, and it
comes back empty rather than missing.
19. Each line under `deviations:` is `builder.add_edge(source, earlier,
condition=fn)` with `fn` named after the key. Cycles are legal and unchecked,
but raise `max_iterations` well above its default of 100, because several live
cycles will exhaust a budget sized for a straight line and fail with a message
about convergence that reads like a broken graph. Keep the ordinary forward
edge unconditional and add each deviation beside it: a condition that returns
false is dropped with no event, so a forward path expressed as a condition
dies silently on every normal run, which is most of them.
20. Handoff values go in `ctx.set_state(key, value)` and come back from
`ctx.get_state(key)`, untyped and unchecked. A write is visible to its writer
at once and to everyone else only in the next superstep, and two writers of one
key in a superstep keep the last write. Never read a key in the same superstep
another phase wrote it.
21. There are no timers, no deadlines and no scheduled wakes. Nothing in
`run-scoped:` becomes an executor and `by:` has no expression at all, so write
them as comments naming where they start and stop, and say plainly in the
fidelity note that a run can close over an unfinished run-scoped line. Do not
fake a deadline with a sleep inside a handler, which blocks the whole superstep
barrier, and never let an expiring wait release a gate.
Produce a second file alongside it, `FIDELITY.md`, and treat it as the more
important of the two. The code is for whoever builds this. The fidelity note
is for whoever has to decide whether this platform suits the process at all,
and that is usually a different person who will never read the code.
It has three parts.
**What came across.** Briefly: how many phases became steps, how many handoff
keys became state fields, which gates stop the run, which deviations became
edges. Counts and names, not reassurance.
**What did not, and what was done instead.** One entry per gap. For each one,
say what the document requires, what the platform can actually express, what
you did in its place, and what breaks if somebody later removes your
workaround. This last part matters most: a workaround nobody understands is a
workaround somebody deletes.
**What a person still has to decide.** The blanks are not a translation
failure, they are the point of a reference process, so list what has to be
filled in before this could run against anything real, and say which of those
choices the platform constrains.
Write it in plain English for somebody who has not read the specification, and
do not soften it. A translation of a reference process is a draft to argue
with, not a build artifact, and the honest account of what was lost is the most
useful thing you will produce.
Here is the process document.
```
PROCESS: instrument a process id: <team>/instrument-a-process v1
from: ref/rev/instrument-a-process v1
owner: <who> effective: <date>
trigger: a process publishes or amends at a version in <your process
register>, or <who> asks for an existing process to be measured
watch: record=<process document>
system=<your process register>
change=<a process publishes or amends at a version>
or watch: record=<the request to measure>
system=<where requests are raised>
change=<somebody asks for an existing process to be
measured>
concurrency: runs may overlap - <how many> processes are being
instrumented at once, and one process is instrumented by
one run at a time
goal: every step the owner named as worth watching produces evidence on
its own, the evidence separates a followed run from a departed
one, and the numbers reach the review on a fixed cadence
phases:
take-in-process - human: read the published process at the version
being measured, and the steps it says are gates
owner: process-librarian
after: trigger
automation: assisted
pick-steps - convenes decide-and-announce: which steps matter
enough to watch. The owner names them, and the
steps left unwatched are written down as
deliberately unwatched
owner: adoption-monitor
after: take-in-process
by: <days>
automation: <level>
define-departure - human: for each watched step, what a followed
step looks like in the data and what a skipped
one looks like, in words a person can check
owner: adoption-monitor
after: pick-steps
automation: never
find-evidence - runs collect-and-report: the fields, timestamps
and logs that already carry the behavior, each
named with the system it sits in
owner: analytics
after: define-departure
by: <days>
automation: supervised
name-gaps - runs assessment: the departures that no system
would show today, and what it would take to see
each one
owner: adoption-monitor
after: find-evidence
automation: <level>
design-changes - human: the fields, events and log lines that
have to change, written as a change request
against <your systems>
owner: adoption-monitor
after: name-gaps
by: <days>
automation: assisted
check-policy - runs assessment: the proposed watching scored
against <your privacy policy> and <your
employment agreements>, per field
owner: standards-keeper
after: design-changes
by: <days>
automation: <level>
approve-change - convenes approval: the systems owner signs the
change, and the privacy reviewer signs the
watching, as named signers
owner: <your systems owner role>
after: check-policy
by: <days>
automation: never
build-instruments - human: the fields, the events and the logging
are made in the systems, at a version, with a
way to turn each one off
owner: <your systems administrator role>
after: approve-change
by: <days>
automation: <level>
build-baseline - runs collect-and-report: what following the
process looked like over the last <how many>
periods, from the evidence that already existed
owner: analytics
after: find-evidence
by: <days>
automation: supervised
test-detection - runs assessment: <how many> runs whose answer is
already known, good and bad, put through the
detection and scored against that answer
owner: adoption-monitor
after: build-instruments + build-baseline
automation: <level>
set-thresholds - convenes decide-and-announce: how much departure
is worth raising, to whom, and how fast. <who>
confirms the numbers
owner: decision-coordinator
after: test-detection
automation: never
turn-on-signal - human: the report goes out on <your cadence> and
the alert path opens, both naming the process
and the version they measure
owner: reporting-manager
after: set-thresholds
automation: <level>
tell-the-measured - convenes briefing: the people doing the work
hear from their own manager what is measured,
what is not, and who sees the numbers
owner: <your process owner role>
after: set-thresholds
by: <days>
automation: never
watch-first-weeks - runs collect-and-report: every signal raised in
the first <how many> weeks, checked by hand
against the run it fired on
owner: adoption-monitor
after: turn-on-signal + tell-the-measured
by: <weeks>
automation: <level>
hand-to-review - human: the measure, its baseline and its known
blind spots handed to the adoption review as its
standing input
owner: reporting-manager
after: watch-first-weeks
automation: <level>
run-scoped:
false-alarms - runs collect-and-report owner: adoption-monitor
every: <cadence>
from: turn-on-signal until: run close
status - runs roll-call owner: adoption-monitor
every: <cadence>
from: pick-steps until: turn-on-signal
handoffs:
take-in-process -> pick-steps [published-process]: the process at its
version, with every step it calls a gate marked as a gate
pick-steps -> define-departure [watched-steps]: the watched steps, and
the unwatched ones with the reason each was left out
define-departure -> find-evidence [departure-signs]: for each watched
step, the observable difference between a followed run and a
departed one
find-evidence -> name-gaps [existing-evidence]: the fields and logs
that already carry the behavior, each naming its system and how far
back it goes
find-evidence -> build-baseline [baseline-fields]: the same fields, so
the baseline is built from the evidence that existed before the
change
name-gaps -> design-changes [evidence-gaps]: the departures nothing
would show, and what each one would cost to see
design-changes -> check-policy [change-request]: the change request,
field by field, saying what each new field records about a person
check-policy -> approve-change [policy-score]: the scored request,
with the policy line behind every refusal
approve-change -> build-instruments [approved-change]: the approved
change at the signed version, and the fields the privacy reviewer
struck out
build-instruments -> test-detection [live-instruments]: the fields and
events live, with a dated test showing each one arriving
build-baseline -> test-detection [baseline]: what compliance looked
like before, so a change in the number is not read as a change in
behavior
test-detection -> set-thresholds [detection-score]: the score against
the known runs, with every miss and every false alarm listed
set-thresholds -> turn-on-signal [thresholds]: the numbers, who is
raised to, and how fast
set-thresholds -> tell-the-measured [what-is-measured]: the same
numbers, in the words the people doing the work will hear them in
turn-on-signal -> watch-first-weeks [live-signal]: the report and the
alert path live, each naming the process version it measures
watch-first-weeks -> hand-to-review [tuned-measure]: the tuned
detection, its false alarm rate, and the steps it still cannot see
deviations:
name-gaps -> define-departure [no-system-shows-departure]: departure
cannot be seen in any system, so what departure looks like is
written again or the step moves to the unwatched list
check-policy -> pick-steps [watching-refused]: the policy check
refuses the watching, so the steps worth watching are picked again
from what is left
approve-change -> design-changes [change-cut-back]: the systems owner
refuses or cuts back the change, so the fields and events are
designed again
test-detection -> design-changes [detection-misses-departures]: the
detection misses departures it was shown, so it is designed again
and the same known runs are put through it
watch-first-weeks -> set-thresholds [signal-fires-wrongly]: the signal
fires on runs that followed the process, so the thresholds are set
again and the runs that fired wrongly are named
watch-first-weeks -> define-departure [step-never-observable]: the
first weeks show a watched step nothing can see, so what departure
looks like for that step is written again
watch-first-weeks -> build-instruments [instrument-broke]: a field is
renamed or a screen retired and the evidence stops arriving, so the
instruments are built again rather than the drop being read as
perfect compliance
bindings:
roster: <who holds each role - the agent claiming adoption-monitor,
the agent claiming analytics, the named systems owner, the
privacy reviewer, and the managers who tell their teams>
systems: the CRM (write), <your systems> (write),
<your logging system> (read), <your reporting tool> (write),
the alert path (trigger), the process register (read)
data: the process document <version>, <your privacy policy>
<version>, <your employment agreements>, the baseline period
policy:
nothing is watched that <your privacy policy> does not allow, and the
check runs before the field is built, not after
the people doing the work are told what is measured before the first
report goes out, and they are told by their own manager
a measure names the process version it measures. A process that amends
sends its measure back through test-detection
a signal that has not been checked against a real run in <how many>
weeks is switched off rather than left firing
the report shows a rate and a count, never a ranked list of named
people, unless <who> asks for names in writing
every instrument can be turned off without a deployment
measures:
cycle time: <target> from the published process to the first report
coverage: <target> share of the watched steps that produce evidence
without anybody being asked
precision: <target> share of raised signals that were real departures
quality gate: no field is built before the policy check passes, and no
report goes out before the people in it were told
```
Take it somewhere
Use this process in CrewAI Flows
close
Paste this into an assistant that can read the web, such as Claude, ChatGPT or Cursor. It reads the specification and the current CrewAI documentation, then writes two files: the flow, and a note on what did not survive the translation. Read the note first. What a runtime cannot express is the part worth arguing about, and this process is a draft to argue with.
copy the prompt
403 lines · the document is inside it, so nothing else is needed
Convert the business process below into a runnable CrewAI Flow:
one Python file with a Pydantic state model, one Flow subclass, @start,
@listen, @router and a durable human feedback provider.
The document is a reference process written to the Agent Processes
specification. Read the specification before you start, because it defines
terms that look ordinary and are not:
https://agentcatalog.com/spec/agent-processes
Sections 6 (the phase graph), 6.5.1 (exception edges), 6.7.1 (deviations),
7 (automation) and 8 (handoffs) are the ones this conversion turns on.
Then read the current documentation for the primitives you will need, rather
than relying on what you remember of the API:
https://docs.crewai.com/en/concepts/flows
Flow, @start, @listen, @router, and_, or_, state, kickoff, plot
https://docs.crewai.com/en/learn/human-feedback-in-flows
@human_feedback, the provider protocol, from_pending and resume
https://docs.crewai.com/en/guides/flows/mastering-flow-state
@persist and what persistence actually promises, which is less than it sounds
WHAT THE DOCUMENT ASKS FOR
These hold wherever the process lands, and they matter more than style.
1. Each phase under `phases:` becomes one step, and keeps its name.
2. `after:` gives the edges. `after: a + b` is a join and waits for BOTH.
Reading it as "either" is the defect the specification calls out by name.
3. Every handoff carries a key in square brackets. Each key becomes one field
on the run's state, named exactly as the key with hyphens turned into
underscores, and the sentence beside it becomes that field's comment. The key
is the stable name; the sentence is prose that may be rewritten.
4. A phase MUST NOT begin before its inbound handoff exists. Where that is
checkable, check it in the step rather than assuming it.
5. `automation: never` is a gate a person signs. The run stops there and does
not continue until a person's decision comes back. Do not turn one into a
notification, a log line, or an automatic transition, whatever the queue
looks like.
6. Each line under `deviations:` is a backward or sideways edge, returning to
the phase named on the right. The key in brackets names it, and that name
belongs in the code.
7. A phase whose `after:` reads like "X or Y, whichever could not finish" is an
exception edge: it is entered when those phases FAIL, not when they succeed.
Do not wire it as an ordinary successor.
8. Anything in angle brackets is a blank the adopting organization fills in.
Leave each one as a named constant at the top of the file with a TODO. Do not
invent a value, a threshold or a date.
9. Record the document's `from:` line at the top of the file, so it says which
reference process and which version it was generated from.
10. Run-scoped lines under `run-scoped:` are work that runs alongside the whole
process rather than at one point in it, and a run may not close while one is
unfinished. Say in the code what you did about them, including if the answer
is that the runtime has nowhere to put them.
HOW THAT LOOKS IN CREWAI FLOWS
11. Build a Flow, not a Crew. A Crew is a team of roles with no graph, no join,
no persistence handle and no gate, and cannot express this document at all.
Each phase is one method on a single `Flow` subclass, keeping its name with
hyphens turned into underscores. A phase that genuinely needs role-based agents
builds its own Crew inside its own method body, and the Flow stays the graph.
12. Declare the state as a Pydantic model bound as the type parameter,
`class NegotiateTheAgreement(Flow[NegotiationState])`, and read and write
`self.state.field`. Never use the untyped dict form: the handoff keys are this
process's memory, and an untyped dict turns a misspelt key into a handoff that
is silently absent. Keep the auto-injected `id` field, which is what every
resume depends on.
13. `after: a` is `@listen(a)`. `after: a + b` is `@listen(and_(a, b))`. Never
write `or_` where the document writes `+`.
14. Check every inbound handoff at the top of the method and raise if one is
missing. `@listen` says when a method may run and says nothing about what is
in hand when it does.
15. The join is the trap, and it is proven rather than theoretical. `and_()`
empties its accumulator the moment it fires, so re-entering BOTH arms works
forever, and re-entering ONE arm after it has fired leaves it holding a single
trigger and waiting for the other for good. The cascade drains, CrewAI prints
that the flow completed, and `kickoff()` returns normally. So for any joining
phase that a deviation can send work back into, do not use `and_()` at all:
make the join a `@router` that both arms trigger, which reads the state fields
and emits its label only when every inbound handoff is present. A router is
re-evaluated against durable state every time and is never suppressed by the
once-fired set.
16. Do not write a phase as `@listen(or_(and_(a, b), "some_label"))`. There is
one accumulator per listener, shared across every branch of its condition and
wiped when any branch satisfies, so the label firing while the join is half
full erases the arm that had already arrived.
17. `automation: never` is `@human_feedback(message=..., provider=...)` stacked
under the method's `@listen`, with a provider whose `request_feedback` raises
`HumanFeedbackPending`. The run then persists, returns that object from
`kickoff()`, and a different process answers later with `from_pending(flow_id,
persistence)` and `resume(text)`. Do not take the default `ConsoleProvider`,
which calls `input()`: that gate exists only while somebody is watching a
terminal, and a run started by a scheduler either hangs or takes an empty
string.
18. Silence must not approve, and the platform's default is that it does. With
`emit=[...]` set, an empty resume collapses to `default_outcome`, or to the
first label when that is unset, with no model consulted and nobody named. Treat
an empty or unrecognised answer as a refusal in your own router. And write the
approver's name and the time onto the state yourself, because
`HumanFeedbackResult` carries the text, the outcome and a timestamp but has no
field for the person, which the specification requires.
19. Each line under `deviations:` is a `@router` named after the key, returning
a label named after the same key, with the target subscribing as
`@listen(or_(normal_trigger, "the_label"))`. Route rather than listen
directly, because a router is re-evaluated on every cycle while a top-level
`or_` listener is suppressed after it first fires.
20. An exception edge is a `try` and `except` around the failing phase's body,
recording the failure on the state and emitting an exception label from a
router. Wiring it as `@listen(or_(x, y))` fires when those phases SUCCEED, so
the clearing phase would run on every healthy run.
21. An unimplemented phase is a method with the document's own sentence as its
docstring and a bare `pass`, which is safe because listeners still fire on a
`None` return. An unimplemented `@router` is not safe: returning `None` emits
no label, every phase below it disappears from the run including the gates, and
the flow reports success. A stub router must return a hard-coded label with a
TODO beside it, or raise.
22. Turn persistence on with `@persist(SQLiteFlowPersistence(...))`, then treat
every phase that performs a real act as something that will run twice.
Persistence saves the state fields and nothing else, so a restart rehydrates
the data and runs the graph again from `@start`: a flow that had already sent a
written refusal sends a second one. Guard each acting phase with a state field
it checks and sets.
23. There are no timers, no deadlines and no cadences, and `run-scoped:` has no
counterpart at all. Put each `by:` value as a named constant, name the
run-scoped lines in the module docstring as unimplemented obligations, and say
in the fidelity note that no deadline in this document is enforced by anything.
These absences produce no diagnostic whatsoever, which is exactly why they have
to be written down.
Produce a second file alongside it, `FIDELITY.md`, and treat it as the more
important of the two. The code is for whoever builds this. The fidelity note
is for whoever has to decide whether this platform suits the process at all,
and that is usually a different person who will never read the code.
It has three parts.
**What came across.** Briefly: how many phases became steps, how many handoff
keys became state fields, which gates stop the run, which deviations became
edges. Counts and names, not reassurance.
**What did not, and what was done instead.** One entry per gap. For each one,
say what the document requires, what the platform can actually express, what
you did in its place, and what breaks if somebody later removes your
workaround. This last part matters most: a workaround nobody understands is a
workaround somebody deletes.
**What a person still has to decide.** The blanks are not a translation
failure, they are the point of a reference process, so list what has to be
filled in before this could run against anything real, and say which of those
choices the platform constrains.
Write it in plain English for somebody who has not read the specification, and
do not soften it. A translation of a reference process is a draft to argue
with, not a build artifact, and the honest account of what was lost is the most
useful thing you will produce.
Here is the process document.
```
PROCESS: instrument a process id: <team>/instrument-a-process v1
from: ref/rev/instrument-a-process v1
owner: <who> effective: <date>
trigger: a process publishes or amends at a version in <your process
register>, or <who> asks for an existing process to be measured
watch: record=<process document>
system=<your process register>
change=<a process publishes or amends at a version>
or watch: record=<the request to measure>
system=<where requests are raised>
change=<somebody asks for an existing process to be
measured>
concurrency: runs may overlap - <how many> processes are being
instrumented at once, and one process is instrumented by
one run at a time
goal: every step the owner named as worth watching produces evidence on
its own, the evidence separates a followed run from a departed
one, and the numbers reach the review on a fixed cadence
phases:
take-in-process - human: read the published process at the version
being measured, and the steps it says are gates
owner: process-librarian
after: trigger
automation: assisted
pick-steps - convenes decide-and-announce: which steps matter
enough to watch. The owner names them, and the
steps left unwatched are written down as
deliberately unwatched
owner: adoption-monitor
after: take-in-process
by: <days>
automation: <level>
define-departure - human: for each watched step, what a followed
step looks like in the data and what a skipped
one looks like, in words a person can check
owner: adoption-monitor
after: pick-steps
automation: never
find-evidence - runs collect-and-report: the fields, timestamps
and logs that already carry the behavior, each
named with the system it sits in
owner: analytics
after: define-departure
by: <days>
automation: supervised
name-gaps - runs assessment: the departures that no system
would show today, and what it would take to see
each one
owner: adoption-monitor
after: find-evidence
automation: <level>
design-changes - human: the fields, events and log lines that
have to change, written as a change request
against <your systems>
owner: adoption-monitor
after: name-gaps
by: <days>
automation: assisted
check-policy - runs assessment: the proposed watching scored
against <your privacy policy> and <your
employment agreements>, per field
owner: standards-keeper
after: design-changes
by: <days>
automation: <level>
approve-change - convenes approval: the systems owner signs the
change, and the privacy reviewer signs the
watching, as named signers
owner: <your systems owner role>
after: check-policy
by: <days>
automation: never
build-instruments - human: the fields, the events and the logging
are made in the systems, at a version, with a
way to turn each one off
owner: <your systems administrator role>
after: approve-change
by: <days>
automation: <level>
build-baseline - runs collect-and-report: what following the
process looked like over the last <how many>
periods, from the evidence that already existed
owner: analytics
after: find-evidence
by: <days>
automation: supervised
test-detection - runs assessment: <how many> runs whose answer is
already known, good and bad, put through the
detection and scored against that answer
owner: adoption-monitor
after: build-instruments + build-baseline
automation: <level>
set-thresholds - convenes decide-and-announce: how much departure
is worth raising, to whom, and how fast. <who>
confirms the numbers
owner: decision-coordinator
after: test-detection
automation: never
turn-on-signal - human: the report goes out on <your cadence> and
the alert path opens, both naming the process
and the version they measure
owner: reporting-manager
after: set-thresholds
automation: <level>
tell-the-measured - convenes briefing: the people doing the work
hear from their own manager what is measured,
what is not, and who sees the numbers
owner: <your process owner role>
after: set-thresholds
by: <days>
automation: never
watch-first-weeks - runs collect-and-report: every signal raised in
the first <how many> weeks, checked by hand
against the run it fired on
owner: adoption-monitor
after: turn-on-signal + tell-the-measured
by: <weeks>
automation: <level>
hand-to-review - human: the measure, its baseline and its known
blind spots handed to the adoption review as its
standing input
owner: reporting-manager
after: watch-first-weeks
automation: <level>
run-scoped:
false-alarms - runs collect-and-report owner: adoption-monitor
every: <cadence>
from: turn-on-signal until: run close
status - runs roll-call owner: adoption-monitor
every: <cadence>
from: pick-steps until: turn-on-signal
handoffs:
take-in-process -> pick-steps [published-process]: the process at its
version, with every step it calls a gate marked as a gate
pick-steps -> define-departure [watched-steps]: the watched steps, and
the unwatched ones with the reason each was left out
define-departure -> find-evidence [departure-signs]: for each watched
step, the observable difference between a followed run and a
departed one
find-evidence -> name-gaps [existing-evidence]: the fields and logs
that already carry the behavior, each naming its system and how far
back it goes
find-evidence -> build-baseline [baseline-fields]: the same fields, so
the baseline is built from the evidence that existed before the
change
name-gaps -> design-changes [evidence-gaps]: the departures nothing
would show, and what each one would cost to see
design-changes -> check-policy [change-request]: the change request,
field by field, saying what each new field records about a person
check-policy -> approve-change [policy-score]: the scored request,
with the policy line behind every refusal
approve-change -> build-instruments [approved-change]: the approved
change at the signed version, and the fields the privacy reviewer
struck out
build-instruments -> test-detection [live-instruments]: the fields and
events live, with a dated test showing each one arriving
build-baseline -> test-detection [baseline]: what compliance looked
like before, so a change in the number is not read as a change in
behavior
test-detection -> set-thresholds [detection-score]: the score against
the known runs, with every miss and every false alarm listed
set-thresholds -> turn-on-signal [thresholds]: the numbers, who is
raised to, and how fast
set-thresholds -> tell-the-measured [what-is-measured]: the same
numbers, in the words the people doing the work will hear them in
turn-on-signal -> watch-first-weeks [live-signal]: the report and the
alert path live, each naming the process version it measures
watch-first-weeks -> hand-to-review [tuned-measure]: the tuned
detection, its false alarm rate, and the steps it still cannot see
deviations:
name-gaps -> define-departure [no-system-shows-departure]: departure
cannot be seen in any system, so what departure looks like is
written again or the step moves to the unwatched list
check-policy -> pick-steps [watching-refused]: the policy check
refuses the watching, so the steps worth watching are picked again
from what is left
approve-change -> design-changes [change-cut-back]: the systems owner
refuses or cuts back the change, so the fields and events are
designed again
test-detection -> design-changes [detection-misses-departures]: the
detection misses departures it was shown, so it is designed again
and the same known runs are put through it
watch-first-weeks -> set-thresholds [signal-fires-wrongly]: the signal
fires on runs that followed the process, so the thresholds are set
again and the runs that fired wrongly are named
watch-first-weeks -> define-departure [step-never-observable]: the
first weeks show a watched step nothing can see, so what departure
looks like for that step is written again
watch-first-weeks -> build-instruments [instrument-broke]: a field is
renamed or a screen retired and the evidence stops arriving, so the
instruments are built again rather than the drop being read as
perfect compliance
bindings:
roster: <who holds each role - the agent claiming adoption-monitor,
the agent claiming analytics, the named systems owner, the
privacy reviewer, and the managers who tell their teams>
systems: the CRM (write), <your systems> (write),
<your logging system> (read), <your reporting tool> (write),
the alert path (trigger), the process register (read)
data: the process document <version>, <your privacy policy>
<version>, <your employment agreements>, the baseline period
policy:
nothing is watched that <your privacy policy> does not allow, and the
check runs before the field is built, not after
the people doing the work are told what is measured before the first
report goes out, and they are told by their own manager
a measure names the process version it measures. A process that amends
sends its measure back through test-detection
a signal that has not been checked against a real run in <how many>
weeks is switched off rather than left firing
the report shows a rate and a count, never a ranked list of named
people, unless <who> asks for names in writing
every instrument can be turned off without a deployment
measures:
cycle time: <target> from the published process to the first report
coverage: <target> share of the watched steps that produce evidence
without anybody being asked
precision: <target> share of raised signals that were real departures
quality gate: no field is built before the policy check passes, and no
report goes out before the people in it were told
```
Take it somewhere
Use this process in Google ADK
close
Paste this into an assistant that can read the web, such as Claude, ChatGPT or Cursor. It reads the specification and the current ADK documentation, then writes two files: the workflow, and a note on what did not survive the translation. Read the note first. What a runtime cannot express is the part worth arguing about, and this process is a draft to argue with.
copy the prompt
394 lines · the document is inside it, so nothing else is needed
Convert the business process below into a runnable Google ADK workflow:
one Python file with a Workflow, nodes, routed edges, a JoinNode,
RequestInput gates and a persisting session service.
The document is a reference process written to the Agent Processes
specification. Read the specification before you start, because it defines
terms that look ordinary and are not:
https://agentcatalog.com/spec/agent-processes
Sections 6 (the phase graph), 6.5.1 (exception edges), 6.7.1 (deviations),
7 (automation) and 8 (handoffs) are the ones this conversion turns on.
Then read the current documentation for the primitives you will need, rather
than relying on what you remember of the API:
https://adk.dev/graphs/routes/
nodes, tuple chains, Event(route=), JoinNode, back-edges
https://adk.dev/graphs/human-input/
RequestInput and the rerun_on_resume handoff
https://adk.dev/runtime/resume/
ResumabilityConfig, resuming by invocation id, at-least-once tools
https://adk.dev/graphs/data-handling/
Event.output against state, and the selector syntax in instructions
WHAT THE DOCUMENT ASKS FOR
These hold wherever the process lands, and they matter more than style.
1. Each phase under `phases:` becomes one step, and keeps its name.
2. `after:` gives the edges. `after: a + b` is a join and waits for BOTH.
Reading it as "either" is the defect the specification calls out by name.
3. Every handoff carries a key in square brackets. Each key becomes one field
on the run's state, named exactly as the key with hyphens turned into
underscores, and the sentence beside it becomes that field's comment. The key
is the stable name; the sentence is prose that may be rewritten.
4. A phase MUST NOT begin before its inbound handoff exists. Where that is
checkable, check it in the step rather than assuming it.
5. `automation: never` is a gate a person signs. The run stops there and does
not continue until a person's decision comes back. Do not turn one into a
notification, a log line, or an automatic transition, whatever the queue
looks like.
6. Each line under `deviations:` is a backward or sideways edge, returning to
the phase named on the right. The key in brackets names it, and that name
belongs in the code.
7. A phase whose `after:` reads like "X or Y, whichever could not finish" is an
exception edge: it is entered when those phases FAIL, not when they succeed.
Do not wire it as an ordinary successor.
8. Anything in angle brackets is a blank the adopting organization fills in.
Leave each one as a named constant at the top of the file with a TODO. Do not
invent a value, a threshold or a date.
9. Record the document's `from:` line at the top of the file, so it says which
reference process and which version it was generated from.
10. Run-scoped lines under `run-scoped:` are work that runs alongside the whole
process rather than at one point in it, and a run may not close while one is
unfinished. Say in the code what you did about them, including if the answer
is that the runtime has nowhere to put them.
HOW THAT LOOKS IN GOOGLE ADK
11. Build a `Workflow` from `google.adk.workflow`, and pin `google-adk>=2.0` in
a comment. Do not use `SequentialAgent`, `ParallelAgent` or `LoopAgent`: they
are deprecated in favour of the graph, and they carry their own defects around
state and control flow. The documentation moved to adk.dev, and anything you
remember about nesting agents rather than drawing a graph is out of date.
12. Each phase is one node keeping its name. Take the node kind from how the
phase resolves rather than from taste: a `human:` or `system:` phase is a plain
Python function node, and a `runs` or `convenes` phase is an `Agent`.
13. `after:` gives the edges, written as tuple chains in `edges=[...]`, and the
trigger is the `"START"` keyword. Take the order only from `after:` lines and
never from the order the phases are listed in.
14. `after: a + b` is a `JoinNode`, and it must be guarded, because this is the
worst trap of any runtime here. The join fires when every static predecessor is
marked COMPLETED, nothing ever un-completes a node, and stored outputs are
never cleared. So after a deviation re-runs one arm, the join fires the instant
that arm finishes and hands the next phase LAST PASS'S value for every arm that
did not re-run. It does not stall, it proceeds with stale data, and nothing
logs. Stamp each arm's output with a pass counter or a content hash, and have
the phase after the join compare the stamps and refuse to run when they
disagree.
15. Each line under `deviations:` is a routed back-edge: a router after the
phase on the left returning `Event(route=...)`, named after the key in
brackets, with one arm going back to the phase on the right and one going
forward. An unconditional cycle raises at construction, which is the one place
this model checks your work. Nothing budgets a routed cycle, so add your own
count and stop rather than looping forever.
16. Give every router an explicit `DEFAULT_ROUTE` arm, and route it to a phase
that stops and asks a person. A route value matching no key writes a log
warning, ends that branch, and lets the run finish reporting success with the
rest of the process never having happened.
17. `automation: never` is a `RequestInput` node of its own, never an `Agent`
asking a question. Decorate it `@node(rerun_on_resume=False)` and yield
`RequestInput(message=..., payload=..., response_schema=...)`, so the run
stops, persists, and delivers the person's answer to the node's successor as
its typed input. A resumed workflow runs its tools at least once, so any
irreversible act needs its own duplicate guard.
18. Make the gates durable or say plainly that they are not. Wrap the graph in
`App(..., resumability_config=ResumabilityConfig(is_resumable=True))` and pass
a persisting session service, never the in-memory one. Note in the file that
the command line and the web UI cannot resume a run, so whoever releases these
gates needs an operator surface that somebody has to write.
19. Every `Agent` in the graph gets `mode="single_turn"` and no `sub_agents`.
A non-empty `sub_agents` list silently adds a transfer tool, and a model that
uses it runs a different agent in this node's place while the graph's outgoing
edge fires on schedule regardless: the topology is honoured perfectly and the
work belongs to somebody else.
20. Model failure as a route, not as an exception. A node that raises does not
propagate: the failure is caught, recorded, and shuts the workflow down without
raising to the caller. So a phase that can fail catches its own failure and
returns `Event(route="could-not-finish")`, and the exception phase hangs off
that arm.
21. Keep every blank as a named module-level constant and never interpolate one
into an `instruction=` string. Angle brackets and curly braces are ADK's own
data selector syntax inside instructions, so a blank pasted verbatim stops
being a blank and becomes a selector.
22. `by:` and `not-before:` have no expression, and `@node(timeout=)` is not
one: it is an in-process wall clock that cancels the node and, because failures
are swallowed, ends the run silently rather than recording a missed deadline.
Nothing in `run-scoped:` has an expression either, and it must not be faked as
an ordinary node, because a node has to be reached and has to finish before
anything downstream starts, which is the opposite of what those lines mean.
Leave both out of the graph and name them in the fidelity note.
Produce a second file alongside it, `FIDELITY.md`, and treat it as the more
important of the two. The code is for whoever builds this. The fidelity note
is for whoever has to decide whether this platform suits the process at all,
and that is usually a different person who will never read the code.
It has three parts.
**What came across.** Briefly: how many phases became steps, how many handoff
keys became state fields, which gates stop the run, which deviations became
edges. Counts and names, not reassurance.
**What did not, and what was done instead.** One entry per gap. For each one,
say what the document requires, what the platform can actually express, what
you did in its place, and what breaks if somebody later removes your
workaround. This last part matters most: a workaround nobody understands is a
workaround somebody deletes.
**What a person still has to decide.** The blanks are not a translation
failure, they are the point of a reference process, so list what has to be
filled in before this could run against anything real, and say which of those
choices the platform constrains.
Write it in plain English for somebody who has not read the specification, and
do not soften it. A translation of a reference process is a draft to argue
with, not a build artifact, and the honest account of what was lost is the most
useful thing you will produce.
Here is the process document.
```
PROCESS: instrument a process id: <team>/instrument-a-process v1
from: ref/rev/instrument-a-process v1
owner: <who> effective: <date>
trigger: a process publishes or amends at a version in <your process
register>, or <who> asks for an existing process to be measured
watch: record=<process document>
system=<your process register>
change=<a process publishes or amends at a version>
or watch: record=<the request to measure>
system=<where requests are raised>
change=<somebody asks for an existing process to be
measured>
concurrency: runs may overlap - <how many> processes are being
instrumented at once, and one process is instrumented by
one run at a time
goal: every step the owner named as worth watching produces evidence on
its own, the evidence separates a followed run from a departed
one, and the numbers reach the review on a fixed cadence
phases:
take-in-process - human: read the published process at the version
being measured, and the steps it says are gates
owner: process-librarian
after: trigger
automation: assisted
pick-steps - convenes decide-and-announce: which steps matter
enough to watch. The owner names them, and the
steps left unwatched are written down as
deliberately unwatched
owner: adoption-monitor
after: take-in-process
by: <days>
automation: <level>
define-departure - human: for each watched step, what a followed
step looks like in the data and what a skipped
one looks like, in words a person can check
owner: adoption-monitor
after: pick-steps
automation: never
find-evidence - runs collect-and-report: the fields, timestamps
and logs that already carry the behavior, each
named with the system it sits in
owner: analytics
after: define-departure
by: <days>
automation: supervised
name-gaps - runs assessment: the departures that no system
would show today, and what it would take to see
each one
owner: adoption-monitor
after: find-evidence
automation: <level>
design-changes - human: the fields, events and log lines that
have to change, written as a change request
against <your systems>
owner: adoption-monitor
after: name-gaps
by: <days>
automation: assisted
check-policy - runs assessment: the proposed watching scored
against <your privacy policy> and <your
employment agreements>, per field
owner: standards-keeper
after: design-changes
by: <days>
automation: <level>
approve-change - convenes approval: the systems owner signs the
change, and the privacy reviewer signs the
watching, as named signers
owner: <your systems owner role>
after: check-policy
by: <days>
automation: never
build-instruments - human: the fields, the events and the logging
are made in the systems, at a version, with a
way to turn each one off
owner: <your systems administrator role>
after: approve-change
by: <days>
automation: <level>
build-baseline - runs collect-and-report: what following the
process looked like over the last <how many>
periods, from the evidence that already existed
owner: analytics
after: find-evidence
by: <days>
automation: supervised
test-detection - runs assessment: <how many> runs whose answer is
already known, good and bad, put through the
detection and scored against that answer
owner: adoption-monitor
after: build-instruments + build-baseline
automation: <level>
set-thresholds - convenes decide-and-announce: how much departure
is worth raising, to whom, and how fast. <who>
confirms the numbers
owner: decision-coordinator
after: test-detection
automation: never
turn-on-signal - human: the report goes out on <your cadence> and
the alert path opens, both naming the process
and the version they measure
owner: reporting-manager
after: set-thresholds
automation: <level>
tell-the-measured - convenes briefing: the people doing the work
hear from their own manager what is measured,
what is not, and who sees the numbers
owner: <your process owner role>
after: set-thresholds
by: <days>
automation: never
watch-first-weeks - runs collect-and-report: every signal raised in
the first <how many> weeks, checked by hand
against the run it fired on
owner: adoption-monitor
after: turn-on-signal + tell-the-measured
by: <weeks>
automation: <level>
hand-to-review - human: the measure, its baseline and its known
blind spots handed to the adoption review as its
standing input
owner: reporting-manager
after: watch-first-weeks
automation: <level>
run-scoped:
false-alarms - runs collect-and-report owner: adoption-monitor
every: <cadence>
from: turn-on-signal until: run close
status - runs roll-call owner: adoption-monitor
every: <cadence>
from: pick-steps until: turn-on-signal
handoffs:
take-in-process -> pick-steps [published-process]: the process at its
version, with every step it calls a gate marked as a gate
pick-steps -> define-departure [watched-steps]: the watched steps, and
the unwatched ones with the reason each was left out
define-departure -> find-evidence [departure-signs]: for each watched
step, the observable difference between a followed run and a
departed one
find-evidence -> name-gaps [existing-evidence]: the fields and logs
that already carry the behavior, each naming its system and how far
back it goes
find-evidence -> build-baseline [baseline-fields]: the same fields, so
the baseline is built from the evidence that existed before the
change
name-gaps -> design-changes [evidence-gaps]: the departures nothing
would show, and what each one would cost to see
design-changes -> check-policy [change-request]: the change request,
field by field, saying what each new field records about a person
check-policy -> approve-change [policy-score]: the scored request,
with the policy line behind every refusal
approve-change -> build-instruments [approved-change]: the approved
change at the signed version, and the fields the privacy reviewer
struck out
build-instruments -> test-detection [live-instruments]: the fields and
events live, with a dated test showing each one arriving
build-baseline -> test-detection [baseline]: what compliance looked
like before, so a change in the number is not read as a change in
behavior
test-detection -> set-thresholds [detection-score]: the score against
the known runs, with every miss and every false alarm listed
set-thresholds -> turn-on-signal [thresholds]: the numbers, who is
raised to, and how fast
set-thresholds -> tell-the-measured [what-is-measured]: the same
numbers, in the words the people doing the work will hear them in
turn-on-signal -> watch-first-weeks [live-signal]: the report and the
alert path live, each naming the process version it measures
watch-first-weeks -> hand-to-review [tuned-measure]: the tuned
detection, its false alarm rate, and the steps it still cannot see
deviations:
name-gaps -> define-departure [no-system-shows-departure]: departure
cannot be seen in any system, so what departure looks like is
written again or the step moves to the unwatched list
check-policy -> pick-steps [watching-refused]: the policy check
refuses the watching, so the steps worth watching are picked again
from what is left
approve-change -> design-changes [change-cut-back]: the systems owner
refuses or cuts back the change, so the fields and events are
designed again
test-detection -> design-changes [detection-misses-departures]: the
detection misses departures it was shown, so it is designed again
and the same known runs are put through it
watch-first-weeks -> set-thresholds [signal-fires-wrongly]: the signal
fires on runs that followed the process, so the thresholds are set
again and the runs that fired wrongly are named
watch-first-weeks -> define-departure [step-never-observable]: the
first weeks show a watched step nothing can see, so what departure
looks like for that step is written again
watch-first-weeks -> build-instruments [instrument-broke]: a field is
renamed or a screen retired and the evidence stops arriving, so the
instruments are built again rather than the drop being read as
perfect compliance
bindings:
roster: <who holds each role - the agent claiming adoption-monitor,
the agent claiming analytics, the named systems owner, the
privacy reviewer, and the managers who tell their teams>
systems: the CRM (write), <your systems> (write),
<your logging system> (read), <your reporting tool> (write),
the alert path (trigger), the process register (read)
data: the process document <version>, <your privacy policy>
<version>, <your employment agreements>, the baseline period
policy:
nothing is watched that <your privacy policy> does not allow, and the
check runs before the field is built, not after
the people doing the work are told what is measured before the first
report goes out, and they are told by their own manager
a measure names the process version it measures. A process that amends
sends its measure back through test-detection
a signal that has not been checked against a real run in <how many>
weeks is switched off rather than left firing
the report shows a rate and a count, never a ranked list of named
people, unless <who> asks for names in writing
every instrument can be turned off without a deployment
measures:
cycle time: <target> from the published process to the first report
coverage: <target> share of the watched steps that produce evidence
without anybody being asked
precision: <target> share of raised signals that were real departures
quality gate: no field is built before the policy check passes, and no
report goes out before the people in it were told
```
One run
A simulation of one run
The activities are on the left, whoever is doing the active one is on the right, and the record of the run builds up as it goes.
This run is built from the same rows as the diagram above: the left column is the activity list, the captions are the activity lines, the cast is the roster, and the labels on the wires are what the handoffs say actually passes.
Adoption
What you fill in
52 blanks to fill. Everything else is the process.
this process
from: ref/rev/instrument-a-process v1
Copy this line into your own document. It never claims this process is running anywhere; it records which draft yours started from, and it is what lets the catalog tell you when this one changes.
The header. Your own id, an owner, and the date it takes effect. One line records where it came from, and that line is what lets the catalog tell you when this reference process changes.
The roster. Which agent takes each activity, and which person takes each of the human ones. The process already names what it needs, so this is a lookup rather than a design exercise.
The numbers. Dates, budgets, cadences, and the targets in the measures block. Nothing here can be a reference value, because a target nobody chose is a target nobody meets.