Week 1 at Gauntlet: An AI Agent on the Same Canvas as Humans

Week 1 Challenge at Gauntlet is over! I just shipped a real-time collaborative whiteboard with an AI agent that lives on the same canvas as human users.

Not a sidebar. Sign Up here https://lnkd.in/gfC4FsZX

Not a chatbot. The AI writes to the same shared document as everyone else -- other collaborators watch its edits appear in real-time, no refresh needed.

The trick that made it practical: a deterministic planner that intercepts structured commands before they ever hit the LLM.

"Create a SWOT analysis" generates 16 tool calls from templates -- zero tokens, sub-second, same result every time. Novel commands fall through to GPT-4o-mini with function calling. You get the reliability of hard-coded logic where it matters and the flexibility of an LLM where it doesn't.

The foundation is a Yjs CRDT shared document. Human edits, AI mutations, and concurrent users all converge automatically. No locking, no merge conflicts, no special AI rendering path. The AI is just another writer on the document.

Built with Next.js, Yjs, Supabase, Konva.js, and GPT-4o-mini. Every command traced with token usage and cost.

The lesson: AI agents aren't useful in isolation. They're useful when they participate in the same system as humans.

I wrote up the full build in Building an AI Agent That Shares a Canvas With Humans.

---

Originally posted on LinkedIn.