Loop engineering

Use this contract for every meaningful autonomous task.

Contract

loop:
  id: LOOP-XXX
  task_id: T-XXX

  goal: one bounded outcome

  requirements:
    - FR-XXX

  acceptance:
    - AC-XXX

  inputs:
    - exact relevant files only

  allowed_actions:
    - read
    - edit scoped files
    - run local tests

  forbidden_without_human_approval:
    - production deployment
    - destructive database migration
    - credential change
    - billing or DNS change
    - irreversible external action

  verify:
    - exact commands or observations

  budget:
    max_attempts: 3

  terminal_states:
    - DONE
    - BLOCKED
    - NEEDS_HUMAN
    - FAILED_SAFE

Runtime

1. Goal — state the exact outcome.
2. Context — read the smallest relevant durable state.
3. Plan — choose the next falsifiable step.
4. Act — make the smallest coherent change.
5. Verify — collect real evidence.
6. Diagnose — if failed, update hypothesis.
7. Retry — only within budget and only with a changed hypothesis/new evidence.
8. Record — update task and durable state.
9. Decide — next task or terminal state.

Anti-patterns

Never:

• loop forever,
• retry identical commands without learning,
• let the same agent self-certify purely by confidence,
• hide failing tests,
• broaden scope silently,
• write "done" without evidence.

Completion evidence

Record:

• command/test,
• result,
• affected acceptance criterion,
• evidence file or output summary,
• known limitations.