CopilotKit released CLiF in early access. CLiF turns user and agent actions into skills, evals, and fine-tuning datasets. It captures both agent traces and user actions via AG-UI. CLiF works with any agent harness and any model. It is fully self-hostable. Skills are proposed as readable files for review and approval. Approved skills reach agents without redeploying. Product trajectories can reduce repetitive tasks by up to 80% in two weeks. Agent trajectories are available today. Product trajectories are in early access. CLiF integrates with existing observability and eval stacks. It supports LangGraph, Mastra, Google ADK, and Microsoft Agent Framework.
What changed
Agents learned only from their own conversations in CopilotKit Intelligence.
What it unlocks
Turning user and agent actions into skills, evals, and fine-tuning datasets without redeploying.
- 80% drop in repetitive tasks within two weeks
- 15 eligible threads minimum per learning run
What you need to act on it
- early access sign-up
- AG-UI protocol integration
- self-hosting or managed cloud setup
Sources