Introduction

The Planner + Agents model provides modularity and scalability. The Planner analyzes user intent, generates a dependency graph of tasks, and schedules Agents. Each Agent runs in an isolated context, performing a single responsibility and reporting results back to the Planner.

1. Planner Responsibilities

• Intent Parsing: Converts CLI or UI commands into structured tasks.

• Graph Generation: Builds a directed acyclic graph (DAG) representing task dependencies.

• Scheduling: Determines execution order (sequential or parallel) based on the DAG and resource availability.

2. Agent Lifecycle

• Initialization: Receives context subset from the Planner (via MCP).

• Execution: Performs its function (e.g., generating code, running audits).

• Reporting: Emits status, output artifacts, and metadata back into the shared context.

3. Isolation & Scalability

• Isolated Workers: Agents run in sandboxed processes or containers to avoid side effects.

• Resource Management: Planner allocates CPU/memory quotas and retries on failure.

4. Integration Points

• MCP Context: Shared storage for intermediate artifacts.

• A2A Hooks: Agents can trigger subordinate Agents (e.g., audit-agent invoking patch-agent).

• Plugin API: Teams can register custom Agents by implementing a simple interface (init, run, report).

Embedded Code Example

// Planner generates a simple task graph
const graph = Planner.buildGraph(['init', 'audit', 'test']);
graph.nodes.forEach(node => console.log(node.id, node.deps));

// Agent interface example
type Agent = {
  name: string;
  init(context: Context): Promise<void>

Visual Description

1. Architecture Diagram:

   • Planner Node: Top-level box parsing commands and emitting DAG.

   • Agent Nodes: Four boxes (scaffold, audit, test, deploy) with arrows from Planner.

   • Context Store: Shared database icon with arrows to/from each Agent.

Conclusion

The Planner + Agents architecture delivers clear separation of concerns, enabling CODI to scale and adapt. Developers benefit from predictable task orchestration and easy extension via custom Agents.