Key points
- CLiF captures full-stack trajectories of both agent behaviour and user actions.
- The platform converts these records into skills, evals and fine-tuning data.
- CopilotKit remains open source with over 38,000 GitHub stars.
Teams shipping AI agents at scale face new tools for improving performance. CopilotKit has released CLiF, a platform designed to improve how agents learn from real-world usage. This early access release aims to solve consistency issues in complex tasks by combining agent traces with actual user behaviour data.
The tool addresses a gap where agent observability often misses the correct workflow path. By recording what agents do alongside what users do, CLiF creates a unified timeline. This allows teams to generate evaluation datasets and skills that reflect actual product usage rather than just theoretical prompts or isolated agent errors.
Technical approach and data capture
Traditional tracing tools show where an agent fails but not how a human would succeed. Product analytics capture clicks and sessions but are not designed for agent training. CLiF merges these views using the AG-UI protocol. It securely captures the full-stack trajectory, linking the agent’s internal logic with the external user actions that represent the correct path for any given task.
This four-stage loop processes the combined data into usable assets. Teams receive skills, production-based evaluations, and fine-tuning datasets. Every skill generated undergoes human review before deployment. The raw trajectories are also exportable, allowing organisations to maintain control over their data while leveraging automated learning processes to refine agent behaviour.
Impact on development teams
Teams deploying AI agents at scale often struggle with hand-written evaluation cases. They may create hundreds of scenarios to test workflows, then adjust prompts and fixes incrementally. CLiF targets the pains slowing agent teams, including evals, context bloat, cost, fragmented stacks, and shadow AI usage. It offers a self-hosted solution that keeps sensitive interaction data within the organisation’s own infrastructure.
The platform supports models like Fable and Astra, which can struggle with end-to-end tasks. By learning from how team leads actually perform tasks, such as closing a sprint, agents gain practical skills. This approach moves beyond static instructions, enabling agents to understand the nuanced sequence of actions required in specific product environments.
Our take on the response
The release of CLiF suggests a shift towards more dynamic agent training methods. By focusing on full-stack trajectories, CopilotKit addresses a fundamental limitation in current agent stacks. This could reduce the manual effort required for evaluation and fine-tuning, potentially accelerating deployment timelines for complex AI applications while maintaining data security through self-hosted options.
What affected people can do
Development teams should review how they currently evaluate agent performance. Check whether your tracing tools capture user actions alongside agent logs. Verify that your data pipelines can merge these distinct data sources effectively. Look for manual workarounds in your evaluation processes that this platform might automate. Ensure human review steps are integrated into any new skill generation workflows. Monitor for changes in agent behaviour after implementing new training data to catch unexpected deviations early.
What to do and how to stay safe: CopilotKit
- Audit your current agent tracing tools to see if they capture user actions.
- Verify that your data pipeline can merge agent logs with product analytics.
- Implement human review processes for any automatically generated agent skills.
- Monitor agent performance closely after integrating new training datasets.
Step-by-step guide: Detect Online Tracking: Signs in Logs, Browsers and Network Traffic
General security guidance from the Breach Observer newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What is CLiF from CopilotKit?
CLiF is a continual learning platform that turns agent traces and user actions into skills, evaluations, and fine-tuning data.
Is CopilotKit open source?
Yes, CopilotKit is open source and has more than 38,000 stars on GitHub.
How does CLiF capture data?
It securely captures full-stack trajectories, including agent traces and user actions, as a single timeline over AG-UI.