Trust and policy

Earned trust, the agent policy file, and who is responsible for what.

#Earned trust

Each action type has a lane: review each change (the default) or auto (earned after a streak of approvals). Some types are always human. Owners can revoke an auto lane; the ledger and the reason are audited.

#The agent policy

The policy is a versioned document that says which tables an agent may read, row caps, which tools are allowed, and which actions need a human. It is validated and enforced on the server, and changes to it are made deliberately by an owner.

#RACI

Agent and human responsibilities share one matrix. Built-in rules: exactly one human is accountable for each activity, an agent is never accountable, and every always-human action has a human responsible or accountable.

#Performance

Settings → Agent & autonomy shows how each agent performs: decision outcomes, edit rate, time to decision, and where its proposals are rejected, so you can tighten or loosen trust with evidence.