An explicit autonomy contract that the user switches on with phrases like “plow ahead”, “do not stop”, or “finish while I am away”. The agent turns routine questions into stated assumptions, picks the smallest reversible option, keeps working through ordinary test failures and ambiguity, and “do[es] not pause merely to ask which reasonable option the user prefers. Pick one, record why, and keep going.”
Stop conditions are listed instead of left to judgment: missing credentials, paid services, or private data; destructive, irreversible, or production-mutating steps; branch operations, history rewrites, force pushes, or deletions the user did not ask for; high legal, safety, privacy, or security risk; decisions the user reserved; and a verification failure that repeats after reasonable investigation, where the next fix would be speculative. When it stops, it leaves a self-contained handoff.
Decision rules for choosing without the user: reuse existing patterns, prefer local and low-blast-radius changes, keep scope tight, pick correctness over cleverness, validate small first, and, when options are close, “choose the one that is easier for the user or a reviewer to understand later.” It keeps a lightweight decision log and ends with a recap under Goal, Key decisions, Changes, Validation, and Remaining risk. The recap is written for a human who was not watching, which is exactly the reader artifact-decision-log was minted for.
Off-stage: this sets how much autonomy the agent has without performing a lifecycle step. equivalent_to: pstack-principle-never-block-on-the-human (proceed on reversible work, pause for irreversible writes).
See Also
- pstack-principle-never-block-on-the-human · pstack-show-me-your-work — the same contract and its audit trail in pstack.
- pattern-autonomous-loop — the pattern this enters.
- builder-quick-recap — the one-line status signal that ends the recap.