ai-command-bus

Open-source command bus templates for small teams running multi-agent AI workflows.

ai-command-bus turns scattered instructions from chat, Notion, GitHub, Telegram, or webhooks into structured work items. It keeps worker outputs in submission queues and routes unsafe actions through human approval instead of letting agents publish, delete, pay, or message customers autonomously.

Why

Multi-agent work often fails at coordination rather than intelligence:

• instructions stay trapped in individual chat sessions
• workers duplicate or miss tasks
• results arrive without a review path
• unsafe actions happen before a human decision
• teams cannot tell which facts are confirmed, assumed, or blocked

This project provides a small, portable operating layer for those problems.

Core Model

intake -> task queue -> dispatch -> worker submission -> review queue -> human decision

Safety Rules

External side effects are approval-gated by default:

• publishing
• deleting
• purchasing or paying
• changing account/security settings
• sending customer-facing messages
• making legal, financial, medical, or tax claims as final advice

Quick Start

npm install
npm test
node src/router.js examples/task_queue.json

Documentation

• Architecture
• Example task queue
• Example worker submission

Example Work Item

{
  "id": "TASK-001",
  "title": "Summarize new issue reports",
  "status": "ready",
  "risk": "low",
  "target": "research",
  "requiresHumanApproval": false
}

Repository Status

This is an early public-friendly extraction from a working internal operations prototype. The public version intentionally avoids private business data, production credentials, personal workspace URLs, and customer workflows.

License

MIT