CopilotKit, the company behind an open protocol for connecting AI agents to app interfaces, has released CLiF, a system that records how people do a task inside a product and uses those records to teach the product’s AI agent the same task. The launch is split. Learning from the agent’s own conversations works now. Learning from what users do on screen, which is the idea the whole launch is built around, is in early access, according to CopilotKit’s blog post of 13 October.
The usual route to a reliable agent is slow. Teams write evals (tests that check whether the agent handles a given situation) by hand, sometimes thousands of them, then patch the agent until enough pass. CopilotKit’s authors say this bloats prompts with edge cases and fills the surrounding code with hardcoded step orders and retries. CLiF starts from a different premise: the people using the product already know how the job is done, so record them.
CopilotKit’s own illustration is a support desk. A customer writes in saying they were billed twice. The rep confirms the duplicate, notices the subscription runs yearly, and hands back store credit rather than cash, because that is the team’s habit. Neither a policy document nor a prompt records this. An agent has no way to guess it, and nobody writes it down until the agent botches a case. The post says there are hundreds of rules like it.
The mechanics are concrete. CLiF captures what it calls a full-stack trajectory over AG-UI, CopilotKit’s open protocol. That pairs two records. One is the agent’s side: the steps it took, the tools it called and what came back. The other is the human’s side: what the screen showed, which buttons got pressed and which requests went to the server. On a schedule, it looks for recurring requests and the friction people quietly work around, then proposes a skill. Each skill is a SKILL.md file that a human reviews and approves. Once approved, a skill goes live with no new deployment, and CopilotKit says each one stays readable and can be rolled back. The same material also yields training data for smaller, cheaper models, tests built from real usage, and exports to other monitoring tools.
Scheduled analysis starts once a container, the group of conversations that learns together, holds at least 15 eligible threads. Delivery adapters exist for CopilotKit’s own agent, Microsoft Agent Framework, Google ADK, Mastra and LangGraph (Python and TypeScript); any other agent can take the skills as plain files. CopilotKit describes AG-UI as supported by Google, Microsoft, Amazon and most of the ecosystem. That is the protocol’s backer characterising third parties, and the adapter list is the firmer evidence of who has actually wired it in.
One number needs care. The post says teams using product trajectories “can see” a drop in repetitive tasks “up to 80%” within two weeks. It names no customer, gives no sample size or method, and never says what counts as repetitive. It is a marketing ceiling from the vendor, not a result, and the post cites no outside evaluation.
The privacy question is built into the design. Capturing every click and server request a person makes, in order to train software, is monitoring of employees and customers whatever the purpose. CopilotKit’s answer is self-hosting: CLiF runs on AWS ECS/Fargate or, via a Helm chart, on Kubernetes, against a Postgres database, a Redis instance and an identity provider (Okta, Azure AD or Keycloak) that you supply, and trajectories stay inside your network. The post says a managed cloud option covers learning from agent conversations; it makes no such statement for product trajectories.
The approval step is the part to test first. If reviewers approve skills the way people click through cookie banners, the lessons the agent learns will be whatever the most common behaviour was, including the bad habits. A team that tries this should count how many proposed skills its reviewers actually reject in the first month.
Based on CopilotKit’s blog post “Introducing CLiF” by Anmol Baranwal and Eli Berman, 13 October 2026.
