CopilotKit Releases CLiF: A Continual Learning Platform That Turns Agent Traces and User Actions into Skills, Evals, and Fine-Tuning Data
MarkTechPost Asif Razzaq
CopilotKit launched CLiF, an early-access system that learns from what agents and users actually do inside a product. It turns those traces into skills, evals, and training data instead of more guesswork.
Based on reporting by MarkTechPost, Asif Razzaq — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
CopilotKit has shipped a new answer to a very familiar AI-agent problem: the agent trace is only half the story. CLiF, now in early access as part of CopilotKit Intelligence, is built to capture what the agent did and what the user did in the product, then turn that combined record into something a team can actually reuse.
That matters because most observability tools stop at prompts, tool calls, and outputs. Useful, sure. But they miss the part where a real user clicks through the product, fixes the agent’s mistake, or does the task in the order that actually works. CopilotKit’s bet is that those actions are the real training signal, and they’re usually sitting outside the agent stack.
CLiF uses AG-UI, CopilotKit’s open protocol, to put the agent’s events and the product’s events on one timeline. From there, its four-stage loop captures sessions into learning containers, looks for repeated patterns and friction, asks a human to approve any proposed skill, and then delivers the result back to the agent. Those skills are written as plain SKILL.md files, versioned, and reversible. The same trajectories can also produce production-based evals, fine-tuning datasets, and raw exports.
The company’s example is a project-management tool with sprint close-out. An agent following the docs would simply move unfinished issues to the next sprint. CLiF, watching real usage, can surface the extra cleanup team leads actually do afterward: moving already-carry-over issues back to the backlog, unassigning them, and commenting on blocked issues. That gets turned into a skill, then into eval cases, and, if needed, into training data for a smaller model.
CopilotKit is also leaning hard into deployment flexibility. CLiF is fully self-hostable, can run inside a company’s own infrastructure, and works with any agent framework through AG-UI. The pitch is blunt: keep the eval and observability tools you already use, but stop asking them to learn from an incomplete record.
My take — AI-written commentary, not fact-checked reporting
This is the right kind of boring. Not another promise that an agent will “learn” from vibes, but a system that records the actual work and makes a human approve the lesson before it ships. The industry has spent years treating agent traces like a full story when they’re really just the first page, which is a very AI thing to do.
Read more about this at: MarkTechPost