Autonomy

Autonomy controls what the engine may do without asking. It never changes what the engine can observe, and it never bypasses project scope, spend limits, or live verification.

Modes

Choose a mode in Project Settings → Autonomy:

• Audit only / MCP — the engine observes and files findings, and can draft plans, but performs no changes. You or your external agent do every action and report back in the task thread.
• Copilot — the engine executes low-risk, reversible work on its own and asks before risky, external, or paid actions. New projects start here.
• Full — the engine executes everything its policy allows, including paid evidence collection within the project budget. It can spend up to a per-action limit without asking.
• Custom — you set the default rule and override it per department or per action class.

The named modes are presets over the same rule set. Every rule decides whether the engine may Create tasks, Prepare plans, and Execute, the Maximum risk it may take, its Spend limit (USD) per prepared action, and whether it must Require approval before executing.

Risk levels

Each action is classified before it runs. From least to most consequential: observe, draft, reversible, material, structural, external, and spend. A rule's maximum risk caps what the engine may execute; anything above it goes to you in Needs you.

Plans and approval

Before executing, the engine freezes an action plan: the target, the intended change, the risk level and reversibility, estimated cost, the steps, and any warnings. The plan appears at the top of the task thread. When approval is required, choose Approve plan or Reject; approving executes that exact plan. If the target or policy changes first, the plan is superseded and must be prepared again.

Ask the assistant in the thread to explain a plan before you decide — what it changes, what the risk and reversibility mean in practice, and what could go wrong.

Recommendation

Start with Copilot. Move a department or action class to a broader rule once its receipts and measured outcomes have earned that trust, and keep external actions such as directory submissions in your hands until you are comfortable with the engine's judgement.