Build around durable organizational logic, not temporary model workarounds.
There should be one canonical owner for each durable definition.
Preserve only what strengthens the reproducible core.
Treat organizations as:
- company
- role
- SOP
- skill
- step
Do not assume each agent needs a different installed skill set.
Prefer:
- one shared skill library
- package-based grouping
- selective activation by role or task
This keeps the system reproducible while preserving clear behavior boundaries.
A knowledge graph should hold promoted, durable knowledge rather than every raw output.