Multi-Agent Swarm Canvas: Orchestrate AI Agent Swarms Visually

A single agent is one context window and one tool loop. That works until a task spans many modules, services, or research threads. A multi-agent swarm splits the work across specialized workers — an explorer to map the codebase, an editor to change it, a verifier to test it — each with its own context and permissions. No single window has to hold everything, so quality and latency both improve. In classic multi-agent orchestration a supervisor decomposes the goal and workers execute in parallel, all coordinated by a deterministic runtime rather than freeform chatter.

Smoke Monkey Canvas turns a swarm into something you can see and steer. It is a visual spatial multi-agent OS: an infinite 2D canvas where each agent is a card you can drag, zoom, connect, and inspect. Launch it in seconds with npx @smoke-monkey/canvas start — no config, no account, no cloud lock-in. Because Canvas is built directly on top of Smoke Monkey Harness, every agent on the canvas inherits the same 6-phase state machine (explore → plan → edit → verify → recover → complete), the same 24 built-in tools, and the same human-in-the-loop pauses. Instead of reading interleaved terminal logs, you watch agents work side by side on one infinite canvas.

The most reliable swarm shape is supervisor + workers. The supervisor reads the repository, emits a typed task plan, and spawns one worker per task; an orchestrator aggregates their structured outputs and the supervisor synthesizes a final answer. On the Canvas this topology is drawn literally: drop a supervisor card, wire it to worker cards, and watch results flow back along the edges. Under the hood it is the ordinary TypeScript API — Canvas is an observability and control surface over the same runtime.

Use a swarm when the task is wide (many files, services, or sources), parallelizable (subtasks do not touch the same files), or risk-sensitive (a verifier agent should check the editor). Stay single-agent when the task is small, sequential, or shares one tight context. A good rule: if a human would hand it to a team, use a swarm; if a human would do it in one sitting, keep it single. To keep large swarms safe, gate destructive tools with permission pauses, give workers isolated subcontexts, and bound retries with the state machine.