Six choices that shape it
- Judgement and mechanics are separate. The agent analyzes, designs, reviews and decides what to remember, guided by skills. Shell scripts keep state, pick the relevant knowledge, run checks, install and clean up — without calling an AI model, so they are testable and cost nothing to run.
- Process follows risk. Trivial and local work finishes without ceremony; architectural and critical work needs an approved design and a stronger review. A change that is small to type but dangerous to undo still takes the heavier route.
- Knowledge has authority and a lifecycle. Durable knowledge is selected by path, domain, topic and stage. What an agent inferred stays proposed, and reaches no other agent until a human accepts it.
- Task notes are temporary. Plans and investigation notes stay in a local, ignored workspace. Only what should outlive the task becomes knowledge; the workspace is cleaned up after the merge.
- The process is installed like code. Skills, scripts and profiles are committed with the project. Upgrades replace what the framework owns and keep what the project changed.
- How far the agent may go is one person’s setting. Whether a finished task is left in the working tree or arrives as a merged pull request, and whether it may run a whole task without stopping to ask, lives in a gitignored file and is enforced by a script — never by the shared project configuration, so no one grants that trust on a colleague’s behalf.