Agentic Search — Policy Iteration
Scenario-aware policies and evidence-driven iteration for search agents.
My current internship at TikTok Search focuses on Agentic Search and Tako.
- I lead Auto-Skill policy iteration: diagnose concrete failures, propose bounded changes, and compare versioned strategies on the real agent chain.
- Scenario-specific prompts address different user intentions while preserving common planning and response contracts.
- Evaluation considers the delivered search experience alongside agent decisions, failure modes, latency, and cost.
- The iteration process separates policy optimization from independent evaluation and supports traceable, reversible changes.
My role centers on policy design and iteration, collaborating with colleagues responsible for evaluation and shared infrastructure.