Rebyte Loop
Turn reviewed Agent work into a better model for your application.
On this page
One loop, clear ownershipStart with your taskKeep every experiment inspectablePreview scopeRebyte Loop connects real Agent work, human feedback, training data and model versions. Run a task, inspect what happened, correct the answer when needed, then train and evaluate a new version. You decide what data to use and when to publish a change.
Group real Agent Sessions under one Scenario. Each Turn records its result, trace and model version.
One loop, clear ownership
| Step | Your application | Rebyte |
|---|---|---|
| Run | Bind Sessions to a Scenario for one business task. | Execute Turns and record their model version and trace. |
| Review | Assess the result or supply a complete corrected answer. | Keep feedback linked to the exact Turn. |
| Select | Choose all eligible data, untrained data, or specific Turns. | Freeze an immutable, inspectable training snapshot. |
| Train | Start a run with a stable retry key and explicit settings. | Prepare the text, train the model and report progress. |
| Publish | Test the candidate on held-out tasks and approve it. | Route new Sessions to the published version. |
Feedback submission never starts training. A successful training run never publishes itself. Existing Sessions keep their original model version after a new version is published.
Start with your task
Use a Scenario such as “order extraction” or “support answer quality.” The Scenario ID stays the same as you change Agent instructions or model versions. Review its Turns, inspect traces, and collect feedback from users, business rules or evaluators.
The current training method is supervised fine-tuning (SFT): learn from good answers and corrected answers. A negative rating without a correction remains available for review, but is excluded from SFT. There is no reinforcement-learning training mode in this preview.
Keep every experiment inspectable
Each training run returns the resolved source,
complete settings, dataset hash and selected examples. Later feedback changes
or deletion do not rewrite that snapshot. all selects current eligible data;
untrained selects examples absent from the chosen source's successful ancestry.
Each output is an immutable model version. Use its version ID for evaluation, then publish it through a stable model family ID. Re-publish an earlier ready version to roll back future Sessions.
Preview scope
Rebyte Loop is in Preview. Feedback and trace features from the Agent API are included here so you can follow the complete improvement process in one place; this label applies to Loop and its evolving training workflow.
Training currently supports Qwen3.6-35B-A3B and text-only captured context. Your organization must have training access. Tool or image context is excluded explicitly rather than silently transformed.
Follow the quickstart for HTTP and TypeScript examples, or continue using the Agent API without training.