| name | agentic-self-distillation |
| description | Use when designing or reviewing self-distillation workflows for agentic models, including trace collection, teacher or judge feedback, rejection sampling, critique, conversion to SFT or preference data, iterative TRL training loops, and safeguards against self-reinforcing errors. |
Agentic Self-Distillation
Use this skill to turn verified agent behavior into better post-training data.
Workflow
- Define the task source and base policy.
- Collect rollouts with full prompts, actions, observations, tool calls,
outputs, and terminal result.
- Verify or judge traces before using them as training data.
- Convert accepted traces into SFT messages, chosen/rejected preference pairs,
or prompt-only tasks with reward metadata.
- Train the next model with TRL and evaluate on held-out tasks.
- Compare against the previous model before promoting the new data recipe.
Guardrails
- Do not distill unverified model outputs directly.
- Keep rejected traces; they are useful for DPO or reward modeling.
- Version teacher model, verifier, prompt, and filter.
- Avoid training on eval tasks or hidden benchmark labels.
- Watch for self-confirming errors where the teacher and judge share the same
blind spot.
References
references/distillation-loop.md: loop design.
references/trace-schema.md: trace fields and conversion targets.