What is in an Agent Plugin package
Every abstract agent in this catalog is published twice: as a document a person reads, and as a package a program can install. The package is generated from the document and from the processes that name the agent, so the two cannot disagree.
The four parts
plugin.json is the manifest: what the job is, which
capabilities it has, and which processes and activities each capability
covers. README.md is the same thing for a person.
skills/ holds one directory per capability, each with a
SKILL.md saying what that capability does and which activities it answers
for. com.agentcatalog.agent/ is the obligations sidecar: the
grants, the records, the escalations and the boundary, carried as data
rather than prose.
A capability is not an activity
Activities are named per process, so the same capability appears under many names. Watching the first weeks of a launch and watching the registrations for a webinar are the same work on a different subject. Grouping them is a judgement a program cannot make from the names, so it is written down and the package is built from it.
That is why a package's skill list is shorter than its activity list, and why the two numbers are both worth reading: the skills say what the agent knows how to do, and the activities say where in real processes that knowledge is called for.
What a package is not
It is not code and it does not run anything. It is a description of a job, shaped so a runtime can read it. What implements the job is yours, and the catalog takes no position on how you build it.