Canvas offers a continual‑improvement platform for AI agents in production, automating the post‑deployment refinement loop. By collecting live traces, scoring behavior with an LLM judge, and proposing validated updates to prompts, tool hooks, stop conditions, and policies, the system iteratively enhances agent performance while reducing manual engineering effort.
Funding
$500K raised to dateRaised to date based on public sources. This may differ from the amount the company actually raised and is based only on what is publicly available on the internet.
Founders
Product
Problem
Enterprises deploying AI agents in production face a costly, manual loop for improving agent performance: engineers must analyze failure modes, modify system prompts, tools, or policies, and re‑evaluate, which slows iteration and limits scalability.
Solution
Canvas provides a continual‑improvement layer that automates the post‑deployment refinement cycle for AI agents. By ingesting live production traces and applying an LLM judge to score agent behavior, the platform identifies systematic failures and proposes targeted updates to the agent’s harness—such as prompt revisions, tool‑use hooks, stop conditions, or policy rules. Each proposed change is validated against a small labeled holdout set, and only improvements that raise holdout accuracy are retained. This iterative, data‑driven loop runs autonomously, continuously tightening the reward signal and reducing reliance on manual engineering effort, thereby increasing agent reliability and performance over time.
Target Audience
Primary customers are enterprises and product teams that operate AI agents in customer‑facing or internal workflow contexts and need a scalable method to continuously improve agent behavior without extensive manual engineering.
Features
- Automated trace collection from running agents and real‑time scoring with an LLM judge
- Harness optimizer that proposes and validates single‑step updates to system prompts, tool hooks, stop conditions, and policy logic
- Small labeled holdout set used for rigorous validation of each harness change
- Persistent filesystem memory storing candidate harnesses, scores, and trace histories for informed iteration
- Open‑source meta‑agent library supporting Claude Agent SDK and extensible to other frameworks
- Demonstrated improvement on benchmark tasks (e.g., 67% → 87% holdout accuracy on tau‑bench airline)