Canvas began as a knowledge work agent.

Our agent operated across GTM and ops: sales, account management, and internal workflows. We deployed it with our enterprise customers and built evals to make improvement measurable.

We watched how our agent performed in real workflows, analyzed failure modes, updated the agent system, and measured whether those changes helped. The process worked, but it was still manual: each improvement cycle required engineers to understand the agent's behavior and modify the system by hand.

We believed that loop itself could be automated. To test this, we built a system that learns from production behavior, proposes targeted updates to the agent system, and applies them. In deployment, the system materially improved agent performance.

We are now generalizing that system. Canvas builds the continual-improvement layer for organizations deploying AI agents in production.