Autonomous Agent Coordination via WebSocket

Tinker

Home Epics Autonomous Agent Coordination via WebSocketEdit

Cancel

Title *

Project *

Description

Enable agents to work autonomously by having Orchestrator coordinate workers and reviewers through WebSocket messaging.

## Overview

Agents (Worker, Reviewer, Orchestrator) run as long-lived sessions connected via WebSocket. Orchestrator receives periodic pings, checks state, and assigns work to available agents.

## Architecture

```
PingJob (every 30s) → checks if Orchestrator idle for 2+ min
                           ↓
                   Pings Orchestrator
                           ↓
           Orchestrator checks available agents
                           ↓
                   If no one available → stop
                           ↓
                   If someone available:
                     - Check pending_audit → assign reviewer
                     - Check failed tickets → notify worker
                     - Check unassigned tickets → assign worker
                           ↓
                   Broadcasts via AgentChannel to agents
```

## Agent States

| Status | Meaning |
|--------|---------|
| `idle` | Available for work |
| `busy` | Actively working |

No `offline` status - WebSocket connects are transient.

## Message Format

| From → To | Message |
|-----------|---------|
| Orchestrator → Worker | `"Work on ticket #58"` |
| Orchestrator → Reviewer | `"Review ticket #58"` |

## Decision Flow

1. Orchestrator receives ping
2. Check: `list_members(role: "worker", status: "idle")` and `list_members(role: "reviewer", status: "idle")`
3. If no one available → return (everyone busy)
4. If someone available:
   - Any `pending_audit` tickets + idle reviewer? → Broadcast to reviewer
   - Any `todo` tickets with assignee (failed review)? → Broadcast to worker
   - Any `todo` tickets without assignee + idle worker? → `assign_ticket()` + broadcast to worker

## Key Principles

- **No event queue** - Orchestrator queries state when pinged
- **Agents don't claim** - Orchestrator assigns, Workers execute
- **Workers mark idle after review submission** - can work on other tickets while review pending
- **No reactive Rails** - just periodic ping job