Multi-Agent Workflows

Learn how to orchestrate multiple AI agents to collaborate on complex development tasks.

Overview

Multi-agent workflows allow you to split complex tasks across multiple AI agents. Each agent runs in its own Pod, and they communicate through Channels and Pod Bindings.

Setting Up a Multi-Agent Workflow

Create a Channel for the task (e.g., 'feature-auth-implementation')
Launch multiple Pods, each with a different role or focus
Link all Pods to the same Channel
Each Pod uses MCP tools to communicate and coordinate
Monitor progress through the Mesh topology view

Scenario: Code Review Workflow

Use two agents — one writes code, another reviews it:

Developer Pod

Writes code based on ticket requirements

Reviewer Pod

Observes the developer's terminal output via binding, reviews changes, and provides feedback through the Channel

The reviewer can use observe_terminal to watch the developer work, then send feedback via the Channel. The developer reads the feedback and iterates.

Scenario: Frontend + Backend Split

Parallelize full-stack development with two specialized agents:

Frontend Pod

Works on UI components, styles, and client logic

Backend Pod

Implements API endpoints, database models, and business logic

Both Pods share a Channel where they coordinate API contracts, data schemas, and integration points. They can use the shared document to maintain an API specification.

Scenario: Test & Development in Parallel

Run testing alongside development for faster iteration:

Development Pod

Implements the feature

Testing Pod

Writes and runs tests for the new code

The testing Pod monitors the development Pod's progress via Channel messages. When the developer signals completion of a module, the tester writes and executes tests, reporting results back through the Channel.

Best Practices

Start with clear task assignments — Give each Pod a focused initial prompt

Use Channels for coordination — Prefer Channel messages over direct terminal access

Link to tickets — Associate Pods with tickets for automatic context

Monitor via Mesh — Use the topology view to track collaboration status

Clean up — Terminate Pods when tasks are complete to free resources