Agents that learn on the job.
unify post-trains your agent’s model on its own production work. Corrections, retries and outcomes become weight updates, gated by your evals.
In the weights, not the prompt.
Most agents are frozen at deployment. They repeat the same mistakes until someone rewrites the prompt or a new base model ships. unify changes the model itself.
- 01CaptureOne SDK call streams your agent’s traces: tool calls, corrections, retries and outcomes.
- 02ScoreWe turn those signals into rewards and verifiers for your domain, so the model knows what better looks like.
- 03TrainReinforcement learning updates the model’s weights on its own work, continually.
- 04ShipEvery checkpoint runs your eval suite and a forgetting check. You approve what goes live.
Blog
Methods and results from post-training agents in production.