Agents in the SDLC

must

AI may assist. AI may not authorize. Agents cannot approve knowledge, merge to protected branches, or waive gates.

May	May not
Read approved claims via the kernel API with a purpose	Scrape this site as policy
Propose patches, tests, threat-model drafts as candidate	Set lifecycle=approved
Use compiled SKILL.md that matches claim versions	Use a stale skill after withdraw
Open a merge request	Self-merge

Tool allowlists and classification ceilings belong on the agent identity, not in the model prompt alone.