Building a Multi-Agent Review Swarm in Notion

The Breakthrough

Ryan Nystrom demonstrates a workflow that offloads entire feature development cycles to an AI agent by integrating remote compute environments directly into Notion, allowing the agent to plan, execute, and self-review code changes without human intervention.

What Actually Worked

• Remote Compute Integration: Instead of using limited sandboxes, the agent connects to dedicated cloud-based coding VMs (nicknamed "Boxy") to explore codebases and execute bash commands.
• Multi-Agent Review Swarm: The author implemented a custom review skill that slices a changeset into domain-specific parts (frontend vs. backend). It then delegates these slices to separate agents tasked with checking for correctness and maintainability.
• Cross-Model Validation: The review swarm runs validation passes using both GPT and Claude 3 Opus, aggregating findings from both models into a single report for the primary agent to action.
• Continuous Loop Execution: The agent is instructed to monitor CI/CD pipelines, automatically fixing type errors or test failures without requiring the engineer to intervene.
• Notion as Source of Truth: The agent maintains a living work journal within a Notion page, logging findings, problems encountered, and requirements, which serves as the context for subsequent development steps.

Context

Software engineers often lose time managing local environments, context switching between project management tools, and manually verifying code. By treating Notion as the central hub for both the initial spoken brief and the final technical documentation, the author eliminates the need to hop between apps. The process relies on heavy upfront investment in defining goals, constraints, and verification standards, which allows the agent to operate autonomously during the build and review phases.

Content References

• tool: Notion AI, Notion, mentioned
• tool: Vercel, mentioned
• tool: Claude 3 Opus, Anthropic, mentioned
• tool: GPT, OpenAI, mentioned