CopilotKit announced the early access release of CLiF, a platform designed for continual learning from full-stack agent trajectories. Building on top of its open-source framework and AG-UI protocol, CLiF captures both agent actions and real user interactions within software applications. By aligning agent execution traces and user session workflows on a single timeline, the platform converts real-world user behavior into production-grade evals, skills, and exportable fine-tuning datasets.

Traditional observability tools typically track agent prompts and tool calls in isolation, while analytics platforms focus on user clicks, leaving a gap in how agents learn context. CLiF solves this via the Agent-User Interaction Protocol (AG-UI), streaming frontend user events and backend agent logs together. Trajectories are scoped into dedicated “learning containers” that automatically analyze data on set schedules to improve domain-specific agent performance.

By leveraging existing user workflows to teach agents optimal task paths, CopilotKit aims to reduce the manual overhead of hand-written evals and prompt engineering. The platform allows engineering teams to replay sessions step-by-step and export raw trajectory data for offline model training.

Why it matters

  • Solves the agent evaluation bottleneck by automatically turning real user interactions into fine-tuning datasets.

  • Unifies agent tracing and product analytics into a single stream using the open AG-UI protocol.

  • Reduces manual prompt engineering and synthetic eval creation by learning directly from production workflows.

Source: marktechpost.com