Review-and-Fix Loops: The Cold Critic, the Fixer, and the Gate Before Full Autonomy

Act II of the agentic coding ladder: the trust rung. We build a review-and-fix loop where one agent critiques a diff while another repairs it, with a human still approving the result. This is wired entirely out of primitives from earlier episodes: subagents, skills, slash commands, hooks, the orchestrator pattern, headless mode, the Agent SDK, git worktrees, and the @claude GitHub Action.

The core idea: a reviewer who wrote the code is the worst reviewer. You want a generator, then a critic in a fresh cold context, then a fixer, then an objective gate.

Concepts and sources:

• Building Effective Agents (evaluator-optimizer, iteration caps)
• Reflexion and Self-Refine
• LLMs Cannot Self-Correct Reasoning Yet (intrinsic self-correction degrades without an external anchor)
• Multi-agent research system (verify high-stakes outputs with a separate pass)

Building it in Claude Code today:

• Custom subagents in the agents folder under dot-claude
• Code Review and security-review
• ultrareview cloud fleet
• Headless mode with json-schema findings
• Agent SDK for the multi-round loop
• GitHub Actions for the hosted version

The pitfall: the fixer reward-hacks the test gate, documented in ImpossibleBench and EvilGenie. Bound it with tool separation, immutable tests, and a PreToolUse hook.

News: Claude Fable 5 lands in Claude Code via v2.1.170, plus v2.1.169 safe mode and the /cd command.