INTRODUCING: CopilotKit CLiF A platform for Continual Learning from Full-stack Trajectories. Teach your agents how work gets done by turning the work users & employees do into: → Exportable datasets (of trajectories) → Evals (bring into platform of choice) → Skills → Fine-tuned models Your users already do the... exact tasks you're trying to teach your agent, every day. Yet months of evals and engineering leave out your most important dataset... Product trajectories. AG-UI already sees every step your agent takes. Now it can also securely capture what your users do: what was on their screen, the buttons they clicked, the server requests they triggered. Together that's a full-stack trajectory. Your people already do the work. Now agents can learn it from them. Compatible with your existing observability and evals stack. Fully self-hostable. CopilotKit🪁 CLiF is in private alpha. Book a demo / Get early access: More below 👇show more

Beyond excited for this! A big move that leverages the full ability of AG-UI

1/ Export full-stack trajectories A full-stack trajectory combines an agent trajectory and a product trajectory into a single unified whole. It's exactly what you need to create a closed, self-improving loop that allows agents to approximate human expertise better over time, providing value at every step. Book a demo and we'll get you started:

2/ Generate Evals from real usage Then export them to your platform of choice. Hand-written evals test what you think users do. CLiF builds them from what users actually did: real tasks, inputs, and edge cases. No one has to imagine the weird ones, because your users already hit them.

3/ Approve suggested Skills Every team has rules nobody wrote down. Say an annual-plan customer gets charged twice. The team issues a credit, not a refund. CLiF finds patterns like this and proposes them as skills. Live today!

4/ Fine-tune models with these datasets Your most important tasks don’t need the biggest model. They need one that’s seen the work done. Trajectories of real tasks are exactly that training set. Use them to fine-tune and RL open models that run your core flows faster, cheaper, and better. The more your product gets used, the better agents perform in it.

We wrote a deep dive article on CLiF if you're interested in learning more about it. Read here:

evals from what users actually did instead of what we think they do.. this one's big

AG-UI has always been the most obvious place for this. It already sits between an agent and its users, so it sees both sides of all the work done in a product.

This is absolutely ground-breaking and I will shout if from the rooftops!

Learning from real trajectories is powerful, and it raises a quiet authority question. If an agent learns from what employees did, does it also inherit what they were allowed to do? Learned behavior should never carry permissions with it. Train the skill, then issue the grant separately.

so my agent learns from me clicking the wrong button 4 times first

This one is huge. I can't wait for the live walkthrough in the webinar that is right around the corner.

@andrewdsouza copilotkit clif turns agent teams' real workflows into skills

Looks great. I think I need to test it.

The hard part with user trajectories as evals is that humans take detours, so a strict step match punishes agents that find a shorter path. Scoring the end state plus a few required checkpoints tends to hold up better than replaying the clicks.