Search Agent Training — Ongoing
Faithful trajectory data, aligned training objectives, and independent evaluation.
I am adapting existing Tako data and training components to Agentic Search. This is ongoing design and integration work.
Data. Preserve the evidence the agent actually observed, its original actions, runtime repairs, and final rendered outcomes. Refinement proposals should retain provenance and use a clear repair owner.
Training. Compare supervised fine-tuning, reinforcement learning, and on-policy distillation according to the behavior each objective should improve. Model-visible inputs and trainable token boundaries require separate validation for each backend.
Evaluation. Assess both native agent behavior and the delivered product: grounded decisions, tool-call validity, usable results, safety, latency, and cost. Use independent cases and regression checks.
Existing infrastructure is reused; measured Search model-training gains are not yet claimed.