Continuous Agent Loop

Patterns for continuous autonomous agent loops with quality gates, evaluations, and recovery controls.

lidge-jun b3b13b8 1.2 KB Updated

File contents

Continuous Agent Loop

Patterns for structuring long-running agent workflows with loop selection, quality gates, and failure recovery.

Loop Selection

Start
  |
  +-- Need strict CI/PR control?           --> continuous-pr loop
  |
  +-- Need task decomposition from spec?   --> DAG-based loop
  |
  +-- Need exploratory parallel generation? --> infinite loop
  |
  +-- default                              --> sequential loop

Recommended Production Stack

  1. Task decomposition from spec or RFC
  2. Quality gates (lint, test, type-check between iterations)
  3. Evaluation harness (automated acceptance checks)
  4. Session persistence (resume after interruption)

Failure Modes

  • Loop churn without measurable progress
  • Repeated retries with the same root cause
  • Merge queue stalls from conflicting changes
  • Cost drift from unbounded model escalation

Recovery

  1. Freeze the loop
  2. Audit recent iterations for root cause
  3. Reduce scope to the failing unit
  4. Replay with explicit acceptance criteria

lidge-jun/cli-jaw-skills/tree/main/continuous-agent-loop commit b3b13b8cdb

Frequently asked questions

npx skillmds@latest add lidge-jun/continuous-agent-loop