Introduction

CODI’s power lies in its modular “Planner + Agents” architecture, underpinned by Multi-Context Protocol (MCP) and Agent-to-Agent (A2A) messaging. This page breaks down each stage of the workflow, illustrating how tasks flow from high-level planning to individual Agent execution and back, ensuring full traceability and collaboration.

1. Task Planning & Context Setup

• Planner Role: Receives developer intent (e.g., init, audit, test, deploy) and generates a task graph.

• MCP Initialization: Creates a root context object containing project metadata, dependencies, and environment variables.

2. Agent Execution Sequence

• Sequential & Parallel Execution: Planner schedules Agents based on graph dependencies—scaffold → audit → test → deploy.

• Execution Sandbox: Each Agent runs in an isolated worker with read/write access to its context nodes.

3. State Persistence via MCP

• Context Tree: MCP persists state at each node (e.g., generated files, audit results, test reports).

• Checkpointing: Automatic snapshots allow rollbacks or replays from any stage.

4. Agent-to-Agent (A2A) Communication

• Messaging Model: Agents exchange structured payloads (JSON) over internal channels.

• Dynamic Handoffs: Agents can spawn sub-agents (e.g., audit-agent invoking patch-agent) and await responses.

5. Visual Workflow Diagram

1. Flowchart: Boxes for Planner, Agents (scaffold, audit, test, deploy), arrows showing MCP and A2A links.

2. Context Graph: Example tree with nodes annotated by Agent output types.

Conclusion

CODI’s Core Workflow ensures that every development action is orchestrated, contextualized, and auditable. By leveraging MCP for state persistence and A2A for dynamic coordination, CODI delivers predictable, repeatable, and secure on-chain development.