Product overview
What is Canvas?
Canvas is an applied research and product lab focused on self-improving AI agents. It builds a continual-improvement layer for agents running in production, covering post-training, evaluation, infrastructure, and observability. The goal is to use production behavior as an input to ongoing agent improvement rather than treating deployment as the end of model development.
Its public work includes MetaAgent, an open-source library that improves agent harnesses from production traces. Canvas also publishes research on reward modeling, evaluator harnesses, continual learning, and the transition from knowledge-work agents to self-improving systems. The available website describes the lab and its research direction, but does not publish a commercial plan or a step-by-step product workflow.
Core Features
- Continual improvement: Builds systems for agents to improve from production experience.
- Post-training: Works on the layer that adapts agent behavior after initial model training.
- Agent evaluations: Develops evaluation and reward-modeling methods for agent systems.
- Production observability: Examines production traces and behavior to inform improvement.
- Agent infrastructure: Creates supporting systems for deploying and maintaining production agents.
Use Cases
- Production agent teams: Improve deployed agents using observed traces and evaluation results.
- Agent research: Study continual learning, reward modeling, and harness optimization.
- Enterprise AI infrastructure: Build an improvement loop around agents operating in real workflows.
Frequently Asked Questions
What does Canvas build?
Canvas builds the post-training and continual-learning layer for production AI agents.
What is MetaAgent?
MetaAgent is an open-source Canvas library that automatically improves agent harnesses from production traces.
Does the website list pricing?
No commercial pricing is presented on the analyzed page.
