OpenFlow
Multi-agent orchestration for coordinated AI workflows with real-time communication and phase management.
OpenFlow
OpenFlow is Codemux's multi-agent orchestration system. It coordinates multiple AI agents working together on a task, with an orchestrator managing delegation, communication, and phase transitions.
Note: OpenFlow is functional in the backend but the React frontend is still being ported. Some features described here may not yet be available in the UI.
How It Works
- Create a run — Define a title, goal, and select which agent roles to use
- Configure agents — For each role, choose a CLI tool (Claude Code, OpenCode), model, and provider
- Start — The orchestrator analyzes the goal, creates a plan, and delegates tasks to agents
- Monitor — Watch agents communicate in real-time via the communication log
- Intervene — Inject messages to redirect agents, provide clarification, or approve phase transitions
Agent Roles
| Role | Purpose |
|---|---|
| Orchestrator | Manages the overall plan and delegates to other agents |
| Planner | Breaks down the goal into actionable steps |
| Builder | Writes code and implements features |
| Reviewer | Reviews code changes for quality and correctness |
| Tester | Writes and runs tests |
| Debugger | Investigates and fixes bugs |
| Researcher | Gathers information and explores approaches |
Communication Protocol
Agents communicate through a structured message log. The orchestrator uses ASSIGN: messages to delegate tasks and tracks completion via DONE: markers. Users can inject messages that are delivered to the orchestrator.
Phase Transitions
OpenFlow runs progress through phases:
- Planning — Orchestrator analyzes the goal and creates a plan
- Active — Agents are working on delegated tasks
- Waiting Approval — Orchestrator pauses for user input
- Completed — All tasks done, orchestrator outputs final summary
Model Compatibility
| CLI + Model | Reliability |
|---|---|
| Claude Code (Claude models) | High — full protocol compliance |
| OpenCode (Claude models) | Good — generally reliable |
| OpenCode (non-Claude models) | Variable — auto-translator helps but some models struggle with the protocol |
Stuck Detection
OpenFlow includes automatic stuck detection:
- Probe (~50s) — Sends a system probe if no agent activity
- Rescue (~60s active, ~90s planning) — Attempts recovery if stuck
User messages wake the orchestration loop immediately, bypassing the normal 5-second polling interval.