The agent plan had every step except where to stop

Short summary

When agents can autonomously merge PRs, plans must explicitly define where they should stop—a problem the author discovered when an agent unexpectedly merged a config-only Renovate change rather than just opening it for review. Agent delegation requires more formal authority boundaries than human delegation because agents complete every step they can justify from the plan text rather than ambiguously interpreting intent. The solution: label each plan slice with explicit authority levels ('Open PR only' vs 'Merge granted') plus rationale, making autonomy handoffs legible before branch protection enforces final gates.

• •Agent plans must explicitly state where the executor should stop (open PR vs. proceed to merge)
• •Plans document what to build and how to verify, but must also define how far the agent can go autonomously
• •Label each slice with permission level and rationale so agent handoffs are legible to reviewers before repository gates enforce them

Generated with AI, which can make mistakes.