CopilotKit released CLiF in early access, a new continual learning platform designed to turn agent traces and user actions into skills, production-based evals, and fine-tuning datasets. Teams shipping AI agents often struggle to make them execute complex, end-to-end tasks consistently, forcing developers to write hundreds of hand-written eval cases and manually refine prompts. Meanwhile, product users perform those exact tasks daily, but traditional stacks fail to capture both agent behavior and user actions on a single timeline.
CLiF solves this gap by capturing full-stack trajectories through the open AG-UI protocol, recording both the agent's tool calls and the user's actual product interactions on the same timeline. The platform utilizes CopilotKit's Automatic Learning loop to process these trajectories into exportable skills, evals, and fine-tuning data, with a human review step for every generated skill. Operating as a self-hosted solution, CLiF aims to address major development bottlenecks such as context bloat, fragmented stacks, and the labor-intensive creation of test evaluations.
Why it matters: This platform bridges the gap between agent observability and user behavior to help developers automate the creation of training data and evaluation cases for AI agents.