Centaur Loop
半人马环 Centaur Loop：面向 AI Agent 反馈闭环、人类治理和记忆复盘的开源工作台 / Human-governed AI feedback loop workbench.
Last verified:
What is Centaur Loop?
Centaur Loop is an open-source workbench for human-governed AI feedback loops that turns agent work into a complete operating cycle. It manages the full business iteration around agent work: plan, approve, execute, review, publish, collect feedback, reflect, remember, and improve the next run. Unlike typical agent systems that end when an answer is generated, Centaur Loop keeps the operational surface open until work is reviewed, measured, and turned into next-cycle guidance.
Key features include an explicit state machine with 10 distinct lifecycle stages (planning, awaiting plan review, generating, awaiting review, awaiting publish, awaiting feedback, reviewing auto, awaiting memory, cycle complete, and repeat with context), human gates where owners approve judgment-heavy tasks, and a system where feedback becomes memory that improves future cycles. The architecture includes loopEngine.ts for state management, loopPlanner.ts for structured planning, loopExecutor.ts for draft generation, loopReviewer.ts for converting feedback into lessons, and adapter boundaries for OpenAI-compatible model access.
Centaur Loop is designed for AI products where feedback matters, particularly for teams building AI systems that need to improve after work leaves chat. It targets developers working on SEO/GEO content growth loops, short-video production cycles, and teams wanting to wrap existing execution systems like LangGraph, Mastra, Temporal, Inngest, or n8n. The project is open-source under MIT license, built with TypeScript, React 18, Vite, and Zustand.
The tool fills the missing layer around agent execution by adding human gates, business outcomes, and next-cycle improvement where cron/schedules only wake jobs, workflow engines only move tasks, and agent frameworks only plan/execute. It makes review, feedback, and memory part of the product surface while keeping publishing accountable and learning from real-world responses.
Centaur Loop pricing
Pricing model: Freemium
Free and open-source under MIT license. No paid tiers mentioned. Users run the workbench locally by cloning the repository and running 'npm install' followed by 'npm run dev'. The demo runtime works without API keys. Users contribute to shape the core package, adapters, storage, notifications, and memory layer.
Centaur Loop pros
- Open-source with MIT license for free use and modification
- Full operating cycle covering plan through memory improvement
- Explicit human gates for taste, compliance, and publishing judgment
- Feedback converts to memory that improves next cycles
- 10-stage state machine for accountable agent work
- OpenAI-compatible runtime works without API keys for demo
- Built with modern stack: TypeScript, React 18, Vite, Zustand
- Small enough to inspect, structured enough to extend
- Core contracts visible: cycle state, checkpoints, memory candidates
- Designed to wrap existing execution systems, not replace them
- Includes implemented SEO/GEO growth loop for content
- Template available for short-video production loop
- Roadmap includes LangGraph, Mastra, Temporal, Inngest integrations
- Manual publishing kept as governed checkpoint for accountability
- Memory candidates let humans choose which lessons become long-term
- Chat messages, cards, and user actions map to runtime state
- Adapter boundaries support tool registry and memory boundaries
Centaur Loop cons
- Still early prototype phase, not production-ready
- Current implementation is just a working React workbench plus TypeScript runtime
- Limited to implemented SEO/GEO loop and short-video template only
- Agent runtime integrations (LangGraph, Mastra, Temporal, n8n) still planned not done
- No cloud hosting - must run workbench locally
- Requires TypeScript/React knowledge to inspect and extend
- Demo runtime works without API keys but production needs model access
- Memory layer, storage, and notifications still need shaping/contribution
Frequently asked questions about Centaur Loop
What is Centaur Loop?
Centaur Loop is the open-source workbench for human-governed AI feedback loops. It turns agent work into a full operating cycle: plan, approve, execute, review, publish, collect feedback, reflect, remember, and improve the next run. It models the whole business iteration around agent work, not just the task execution step.
How is Centaur Loop different from a scheduler or workflow engine?
Centaur Loop is not another scheduler or workflow canvas. While cron/schedules only wake a job at a time and workflow engines only move tasks through steps, Centaur Loop governs the full cycle before and after the job runs. It makes review, feedback, and memory part of the product surface and adds human gates, business outcomes, and next-cycle improvement.
What are human gates in Centaur Loop?
Human gates are explicit checkpoints where people own taste, compliance, publishing, and final judgment. AI runs where it can, then stops at checkpoints for plans, drafts, publishing, feedback, and memory candidates. These become first-class checkpoints in the state machine.
How does feedback become memory?
Real outcomes, screenshots, notes, and metrics are reviewed into memory candidates during the reviewing auto stage. The loopReviewer.ts converts feedback into retrospectives, lessons, and next-cycle suggestions. Humans then decide which lessons deserve long-term memory at the awaiting memory stage.
What technology stack does Centaur Loop use?
Centaur Loop is built with TypeScript, React 18, Vite, and Zustand. It uses an OpenAI-compatible runtime. The core components include loopEngine.ts, loopPlanner.ts, loopExecutor.ts, loopReviewer.ts, loopChat.ts, and an adapters/ directory for model access and tool registry.
What lifecycle stages does the state machine include?
The state machine has 10 stages: (1) planning - generate structured plan from goals and memory, (2) awaiting plan review - pause for owner confirmation, (3) generating - create drafts through AI tools, (4) awaiting review - approve/reject drafts, (5) awaiting publish - manual publishing checkpoint, (6) awaiting feedback - capture metrics and outcomes, (7) reviewing auto - analyze what worked/failed, (8) awaiting memory - select lessons for long-term memory, (9) cycle complete - carry suggestions forward, (10) repeat with context - start again with reviewed experience.
Who is Centaur Loop for?
Centaur Loop is designed for AI products where feedback matters. It targets developers building AI systems that improve after work leaves chat, teams working on SEO/GEO content growth, short-video production with feedback-driven improvement, and anyone wanting to wrap existing execution systems like LangGraph, Mastra, Temporal, Inngest, or n8n-style approvals.
What starter loops are available?
Two starter loops exist: an implemented SEO/GEO Growth Loop for weekly content growth cycles including WeChat articles, Xiaohongshu notes, and SEO/GEO content with plans, drafts, publishing, feedback, review, and memory; and a Short-Video Production Loop template for daily topic and script cycles with topic confirmation, script review, and outcome feedback.
How do I run Centaur Loop locally?
Centaur Loop is open-source under MIT license. Clone the repository and run 'npm install' followed by 'npm run dev'. The demo runtime works without API keys. Users can inspect the state machine and contribute to shape the core package, adapters, storage, notifications, and memory layer.
What integrations are planned for the roadmap?
The roadmap includes Agent Runtime Integrations designed to wrap existing execution systems instead of replacing them. Planned integrations include LangGraph and Mastra, Temporal and Inngest, and n8n-style approvals. The tool is designed to wrap existing workflow and agent runtimes with examples planned for common systems.