Implemented
SEO / GEO Growth Loop
Weekly content growth cycles for search and AI-answer visibility, including plans, drafts, publishing, feedback, review, and memory.
- WeChat articles
- Xiaohongshu notes
- SEO and GEO content
Open-source control plane / AI feedback loops
The workbench for human-governed AI loops.
Centaur Loop turns agent work into a full operating cycle: plan, approve, execute, review, publish, collect feedback, reflect, remember, and improve the next run.
- TypeScript
- React 18
- Vite
- Zustand
- OpenAI-compatible runtime
cycle://content-growth/weekly
Planning
AI proposes the cycle plan
Goal, memories, tool registry, business context, and previous suggestions become structured tasks.
Human Gate
Owner approves what needs judgment
Plans, drafts, publishing, feedback, and memory candidates become first-class checkpoints.
Execution
Agent work continues automatically
Drafts are generated through adapter-ready tools without losing the loop state.
Review
Outcomes become memory
Feedback is reviewed into lessons, corrections, preferences, and next-cycle suggestions.
npm install
npm run dev
# demo runtime works without API keys
npm run dev
# demo runtime works without API keys
What's different
Not another scheduler. Not another workflow canvas.
The loop is the product
Centaur Loop models the whole business iteration around agent work, not just the task execution step.
Human gates are explicit
AI runs where it can, then stops at checkpoints where people own taste, compliance, publishing, and final judgment.
Feedback becomes memory
Real outcomes, screenshots, notes, and metrics are reviewed into memory candidates that improve the next cycle.
Lifecycle
A state machine built for accountable agent work.
Most agent systems end when the answer is generated. Centaur Loop keeps the operational surface open until the work is reviewed, measured, and turned into next-cycle guidance.
- planning
Generate a structured plan and task list from goals and memory.
- awaiting plan review
Pause for owner confirmation or edits before work begins.
- generating
Create drafts through registered AI tool definitions.
- awaiting review
Approve, reject, or request changes per draft.
- awaiting publish
Keep manual publishing visible as a governed checkpoint.
- awaiting feedback
Capture metrics, notes, screenshots, and outcome signals.
- reviewing auto
Analyze what worked, what failed, and why.
- awaiting memory
Ask humans which lessons deserve long-term memory.
- cycle complete
Carry suggestions forward into the next run.
- repeat with context
The loop starts again with reviewed experience.
Architecture
Small enough to inspect. Structured enough to extend.
The current prototype is a working React workbench plus TypeScript loop runtime. It is early, but the core contracts are already visible: cycle state, human checkpoints, feedback, reviews, memory candidates, and adapter boundaries.
Plan -> Approve -> Execute -> Review -> Publish -> Feedback -> Reflect -> Remember -> Next Cycle
loopEngine.ts
Explicit state machine that advances cycles and stops at human gates.
loopPlanner.ts
Turns goals, memory, business context, and tools into a structured plan.
loopExecutor.ts
Generates reviewable drafts and keeps failures inside the cycle record.
loopReviewer.ts
Converts feedback into retrospectives, lessons, and next-cycle suggestions.
loopChat.ts
Maps runtime state to chat messages, cards, and user actions.
adapters/*
OpenAI-compatible model access, tool registry, and memory boundaries.
Starter loops
Designed for AI products where feedback matters.
Template
Short-Video Production Loop
Daily topic and script cycles with feedback-driven improvement and a path toward fast-loop / slow-loop orchestration.
- Topic confirmation
- Script review
- Outcome feedback
Roadmap
Agent Runtime Integrations
Designed to wrap existing execution systems instead of replacing them, with examples planned for common workflow and agent runtimes.
- LangGraph and Mastra
- Temporal and Inngest
- n8n-style approvals
Positioning
The missing layer around agent execution.
| Layer | What it handles | What Centaur Loop adds |
|---|---|---|
| Cron / schedules | Wake a job at a time | Govern the full cycle before and after the job runs. |
| Workflow engines | Move tasks through steps | Make review, feedback, and memory part of the product surface. |
| Agent frameworks | Plan and execute with models/tools | Add human gates, business outcomes, and next-cycle improvement. |
| Publishing bots | Push generated output somewhere | Keep publishing accountable and learn from real-world response. |
Open source / MIT license
Build AI systems that improve after the work leaves chat.
Run the workbench locally, inspect the state machine, and help shape the core package, adapters, storage, notifications, and memory layer.