Building a Multi-Agent Coding Workflow in the IDE

The Planner-Implementer-Verifier Loop

Instead of relying on a single model to both write and validate code, use a multi-agent workflow to enforce feature parity against project specifications. By connecting multiple agents through an Agent Communication Protocol (ACP) within an IDE like PyCharm, you can assign distinct roles to different models to reduce hallucinations and ensure requirements are met.

Implementation Steps

• Audit Phase: Use a high-reasoning model (e.g., OpenAI o1 or similar "soul" models) with high reasoning effort settings to audit existing code against a build.md specification file. Enable full IDE context access to allow the agent to inspect the entire project structure.
• Planning Phase: Instruct the auditor to generate a P0 and P1 implementation plan. Review the provided diffs within the IDE to confirm the proposed changes align with the project goals before execution.
• Implementation Phase: Switch to a different, cost-effective model (e.g., Kimi K3) to execute the implementation plan. Using a different model family for implementation than the one used for auditing helps mitigate shared blind spots.
• Verification Loop: Return to the original auditor model to verify the new implementation. If the verifier identifies gaps, feed that feedback back into the implementer agent for iterative refinement until the code passes the specification audit.

Why This Matters

CLI-based agents often lack the deep IDE context required to understand complex project dependencies. By centralizing multiple agents in an IDE, you gain the ability to verify implementation correctness against requirements, which is critical for large-scale builds where a single agent might claim a task is complete while missing specific feature requirements like particle animations or complex logic.