Introduction

Self-correcting agent chains for Claude Code and opencode.

A chain of agents takes turns. One plans, one codes, one reviews, one tests, one commits. If an agent fails, another agent fixes it and the chain retries — automatically.

You watch. Approve the PR. Done.

How it works

flowchart LR
    A[Agent runs] --> B[PostToolUse hook]
    B --> C[Read chain<br/>decide + routes]
    C --> D{{Haiku picks<br/>next agent}}
    D --> E[MANDATORY<br/>NEXT STEP]
    E --> F[PreToolUse hook<br/>injects prev output]
    F --> A
    D -- route: END --> G([Done])

That's the whole engine. Each lap of the loop = one agent step.

• You define the chain (flow: in YAML — agents and their routes).
• Hooks fire automatically — no glue code.
• Haiku reads the agent's output against the decide: prompt and picks the next route.
• Context carries forward — previous agent's full output is injected into the next agent's prompt.
• Failure is just another route — point it at a debugger agent and the chain self-corrects.

A run, end-to-end

/code implement login feature

[Chain runs automatically — you don't touch anything]
→ planner: creates plan
→ plan-reviewer: approves ✓
→ implementer: writes code
→ code-reviewer: 2 issues → auto-routes back to implementer
→ implementer: fixes
→ code-reviewer: passes ✓
→ tester: FAILS → auto-routes to debugger
→ debugger: fixes
→ tester: PASSES ✓
→ git-manager: commits + creates PR

[You only approve the final PR]

Why Brainbrew?

What it adds to Claude Code

Claude Code has agents and teams. Brainbrew adds the orchestration layer:

	Vanilla Claude Code	With Brainbrew
Agent chaining	Manual (you decide next step)	Automatic (Haiku routing)
Failure recovery	None (you see error, you fix)	Built-in (debugger → retry)
Quality gates	None	Haiku QA + auto-retry
Parallel agents	Teams (manual trigger)	Teams (auto in chain)
Inter-agent state	None	Memory Bus

vs. LangChain / CrewAI

Same orchestration power. Different tradeoffs:

	Brainbrew	LangChain/CrewAI
Config format	YAML + Markdown	Python code
Cost model	Your CC subscription	Per-token API billing
Learning curve	Pick template, start working	Learn framework, write code
Runs inside	Claude Code / opencode	Standalone runtime

10 templates. Ready to run.

No blank-page problem. Pick a template:

• develop — plan → review → implement → parallel(code-quality + security) → test → [fix if fail] → commit
• devops — scan → security → test → deploy → monitor → [rollback if alert]
• marketing — research → write → edit → SEO → publish → analyze
• research — gather → analyze → synthesize → report
• And 6 more (docs, support, data, moderation, review, skill-dev)

Declared in YAML. Your agents stay with you.

Chain config lives in git. Agents are markdown files in .claude/agents/. No vendor lock-in on your components — only the routing engine is brainbrew-specific.

Features

• Self-correcting pipelines — failures auto-route to fixers, then re-enter the chain
• AI-powered routing — Haiku analyzes output and picks the next step
• Agent teams — parallel execution with coordinated synthesis
• Quality gates — subagent-stop hook validates output, retries up to 2x
• Memory Bus — inter-agent state sharing across pipeline runs
• Loop detection — prevents infinite cycles (MAX_AGENT_LOOPS=3)