Budget for feedback loops
AGENT-BUDGET-FOR-FEEDBACK-LOOPS
Summary
Reserve enough time and tokens for checks, failure inspection, and handoff proof. Planning for the feedback loop keeps validation from being squeezed out after the first edit.
Rule
Budget tokens and time for feedback loops.
Why
Agent work needs room for reading, editing, running checks, inspecting failures, and reporting proof. If the task budget only covers the first implementation attempt, the agent is pushed toward optimistic handoff instead of using feedback from tests, docs builds, screenshots, or reviewer notes.
Helps
- Prevents premature handoff and makes validation failures part of the planned work instead of a surprise.
Limits
Do not reserve heavy feedback loops for tiny edits that can be verified directly. Match the loop to the risk, surface area, and cost of being wrong.
Respect the user's budget. Stop after the required checks pass unless a new finding, changed artifact, or unresolved contract justifies more work. Repeated equivalent reviews and status narration consume capacity without adding evidence.
Agent instruction
Reserve time and tokens for validation and handoff. Give each extra pass a distinct question. Stop after required gates pass unless new evidence warrants another pass; budgets are ceilings.Mechanisms
Supported by explicit task budgets, staged plans, validation checklists, long-running command timeouts, and handoffs that separate run checks from skipped checks.