Build AI skills from user behavior with CopilotKit's self-hosted CLiF platform

CopilotKit released CLiF, a platform for continual learning from full-stack trajectories, in early access. The tool aims to help AI agents perform complex tasks by learning from actual user behavior within products, rather than relying solely on static prompts or manual evaluations.

What happened

Teams deploying AI agents at scale often struggle to make them perform complex end-to-end tasks consistently, according to MarkTechPost. Even agents built on top-tier models like Fable and Astra can fail to meet deployment standards for specific products. Developers typically respond by writing hundreds or thousands of hand-crafted evaluations and tweaking prompts until the agent passes enough tests. This process is slow and does not always capture the nuances of real-world usage.

CopilotKit, the maker of the open-source CopilotKit framework, introduced CLiF to address this gap. The platform securely captures both what the agent does and what users do in the product, combining these records into a single timeline via the AG-UI protocol. This allows the system to turn raw activity into actionable training data.

The platform uses a four-stage loop based on CopilotKit’s Automatic Learning to process these full-stack trajectories. It generates skills, production-based evaluations, fine-tuning datasets, and raw exports. Every skill generated by the system undergoes human review before being finalized. The platform is designed to run fully self-hosted to address concerns around data privacy and fragmented technology stacks.

Why it matters

AI agents often fail because they lack context about how real users solve problems. Traditional observability tools track agent errors but miss the correct solution path, which lives in user behavior. By merging agent traces with user actions, teams can teach agents the actual sequence of steps required for complex tasks. This reduces the need for manual prompt engineering and helps agents learn from production data rather than theoretical scenarios.

The fix

• Monitor agent performance against actual user workflows to identify gaps.
• Ensure full-stack trajectories capture both agent actions and user inputs.
• Review automated skills and evaluations before deploying them to production.
• Verify that data capture complies with privacy and security standards.

Why it matters

Integrating user behavior data into agent training loops helps bridge the gap between theoretical capabilities and practical application. Teams usually need to align agent logic with actual product usage patterns to reduce errors. Checking whether observability tools capture the complete task lifecycle ensures agents learn from correct examples. This approach minimizes reliance on static prompts and allows systems to adapt to real-world complexity.

What to do and how to stay safe: CopilotKit

• Audit current agent evaluation methods to see if they reflect real user tasks.
• Check if your observability stack captures both agent traces and user actions.
• Plan for human review of any AI-generated skills or training data.
• Verify that data collection methods meet your organization's privacy policies.

Step-by-step guide: What Does End-to-End Encryption Mean on Messenger: The Real Privacy Impact

General security guidance from the Crypt Weekly newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

What is CopilotKit CLiF?

CLiF is a continual learning platform that turns agent traces and user actions into skills, evaluations, and fine-tuning data.

Is CLiF available for public use?

The platform is currently in early access, according to MarkTechPost.

How does CLiF capture data?

It securely captures agent actions and user product actions as one timeline over the AG-UI protocol.