Editorial illustration for CopilotKit Releases CLiF to Convert Agent Traces Into Fine-Tuning Data
CopilotKit CLiF Converts Agent Traces to Fine-Tuning Data
CopilotKit has a fix for a problem that's been slowing down every team trying to ship AI agents past the demo stage: the agent passes its evals, then falls apart on real users inside the actual product. The company, which built the open-source CopilotKit framework and the AG-UI protocol and has racked up more than 38,000 stars on GitHub, released CLiF in early access this week. The name stands for continual learning from full-stack trajectories, and the pitch is straightforward.
Most teams write hundreds or thousands of evals by hand, then patch prompts and harness logic until the agent clears enough of them to go live. That process misses something sitting in plain sight: the product's own users already know the correct sequence of steps, because they do the task every day. CLiF captures both sides of that gap, the agent's trace and the user's actions in the product, as a single secure timeline.
What CopilotKit does with that combined record, and why tracing alone has never been enough to catch these failures, is where the platform's four-stage loop comes in.
CopilotKit, the company behind the open-source CopilotKit framework and the AG-UI protocol, released CLiF, a platform for continual learning from full-stack trajectories, in early access. CLiF securely captures what the agent does and what users do in the product, and turns that record into skills, production-based evals, and fine-tuning datasets, all exportable along with the raw trajectories.
Why this matters
CLiF is a bet that the real bottleneck in shipping agents isn't model quality, it's the grunt work of turning messy production traces into usable training signal. For teams drowning in hand-written evals, a tool that replays sessions step by step and automatically spins off skills, evals, and fine-tuning data could cut months off the iterate-until-it-passes grind CopilotKit describes. That's worth paying attention to if you're a founder trying to get an agent past demo stage into something that holds up across thousands of real user sessions.
Still, worth watching closely rather than taking at face value. Converting raw trajectories into clean training data at scale is genuinely hard, and "turns traces into skills" can mean anything from a tidy abstraction layer to brittle pattern-matching that falls apart on edge cases. We'd want to see what happens when agents built on different stacks, not just Fable and Astra, feed into this loop, and how much human curation still hides behind the automation.
The idea of continual learning from production is right. Whether CLiF actually delivers it reliably is the open question.
Related Reading
AI Tools & Apps Claude gains shared context in Excel, PowerPoint; Microsoft adds Copilot Cowork
AI Tools & Apps OpenAI Developer: AI Agent "Swarms" Waste Tokens, Create "Coordination Tax
AI Tools & Apps LG's recent webOS update adds Microsoft Copilot app, now removable
From Archives K2K explores NVIDIA AI factories for Liverpool City Region public services
From Archives OpenAI employee allegedly used chatbot to plant recording device at office
Common Questions Answered
What problem does CLiF solve for AI agent development teams?
CLiF addresses the critical issue where AI agents pass their evaluations during testing but fail when deployed to real users in production environments. By converting agent traces into fine-tuning data, CLiF helps teams move past the demo stage and create agents that perform reliably in actual products rather than just in controlled testing scenarios.
What does CLiF stand for and what does it capture?
CLiF stands for continual learning from full-stack trajectories. The platform securely captures both what the agent does and what users do within the product, then transforms that recorded data into reusable skills, production-based evaluations, and fine-tuning datasets that can be exported along with raw trajectories.
How does CLiF reduce development time for shipping AI agents?
CLiF eliminates the time-consuming manual work of writing evaluations and converting messy production traces into usable training signals. By automatically replaying sessions step by step and generating skills, evaluations, and fine-tuning data, the platform can cut months off the iterative development cycle that teams typically experience when trying to get agents production-ready.
What is CopilotKit's track record in the open-source community?
CopilotKit has established itself as a significant player in the AI development space, having built the open-source CopilotKit framework and the AG-UI protocol. The company has accumulated more than 38,000 stars on GitHub, demonstrating substantial community adoption and trust in its tools.
Further Reading
- Evaluating AI agents: moving past the vibe check - Google Developers Blog
- Agent reliability: how to measure and improve AI agents in production - Arize AI
- Automated structural testing of LLM-based agents: methods, framework, and case studies - arXiv
- Execution-First Synthetic Tool-Use Trace Generation for LLM Agents - arXiv
- Introducing CLiF: Continual Learning from Full-Stack Trajectories - CopilotKit
