CopilotKit Releases CLiF for Continual Learning from Full-Stack Trajectories

CopilotKit has released CLiF (Continual Learning from Full-Stack Trajectories) in early access, a tool designed to turn existing user and agent interactions into production-ready skills, evals, and fine-tuning data. By capturing full-stack trajectories—which include both the agent's internal trace and the user's actual actions on the screen—CLiF allows agents to learn from how work is actually done rather than relying solely on manual prompt engineering.

How Full-Stack Trajectories Work

CLiF operates on the AG-UI (Agent-User Interaction Protocol), an open standard supported by Google, Microsoft, and Amazon. This protocol sits between the agent and the user, capturing two distinct data streams:

• Agent trajectory: Every run, tool call, state change, and return value.
• Product trajectory: User clicks, screen states, and server requests triggered by human actions.

By analyzing these combined streams, CLiF identifies patterns and friction points that users silently work around. These insights are converted into proposed skills (formatted as SKILL.md files) which, once approved by a human, are deployed to the agent without requiring a redeploy.

Key Features and Outputs

The system generates several artifacts to improve agent performance:

• Skills: Reusable instructions derived from successful interaction patterns.
• Production-based evals: Test cases built from real user workflows, compatible with existing evaluation stacks.
• Fine-tuning data: Datasets for training smaller, more efficient models on specific tasks.

CLiF supports integration with various agent harnesses, including LangGraph, Mastra, Google ADK, and Microsoft Agent Framework. It allows teams to import existing conversation threads to avoid starting from zero.

Deployment and Availability

CLiF is part of CopilotKit Intelligence and offers two deployment options:

• Managed Cloud: For the fastest setup.
• Self-Hosted: Fully self-hostable, allowing users to bring their own Postgres, Redis, and identity providers (Okta, Azure AD). All data, including trajectories and secrets, remains within the user's VPC or on-premises environment.

While learning from agent trajectories is available now, the product trajectories feature (capturing user actions) is currently in early access. CopilotKit reports that teams using this approach can reduce manual prompt maintenance and improve agent success rates by learning directly from production usage.