Lakshya Agrawal

Lakshya A. Agrawal was described in his 2026 AI Engineer conference biography as the creator and maintainer of GEPA and a second-year EECS PhD student at UC Berkeley’s Sky Computing Lab. His research focuses on optimization, evaluation, and self-improvement for LLM-based agents and systems. He previously worked as an AI4Code Research Fellow at Microsoft Research.

Agrawal presents reflective optimization as a way to improve AI systems by learning from full execution traces, including intermediate outputs, tool responses, and error messages. GEPA uses this feedback to propose better prompts and maintains a Pareto pool of candidates that perform well on individual training examples, helping the search avoid local optima. He also presents Optimize Anything, which extends this approach to code, agent harnesses, scheduling policies, and other text artifacts that can be scored, with evaluators supplying diagnostic feedback alongside results.

His examples include evolving a four-line program into a six-step agent that hypothesizes rules, synthesizes and executes code, and debugs its solutions. He reports that this raised Gemini Flash accuracy on an ARC-AGI example from 32.5% to 89.5%. He also describes learning repository-specific coding skills that improved a GPT-5 mini agent’s Go issue-resolution results from 24% to 93%, with those skills transferring to Claude Sonnet. For subjective tasks, he explains how detailed human feedback on production trajectories can train an LLM judge that guides further agent improvement, and discusses jointly optimizing model weights and prompt harnesses.