Lakshya A Agrawal

I am a Computer Science PhD student at UC Berkeley, advised by Prof. Matei Zaharia and Prof. Dan Klein. I am affiliated with Sky Lab, BAIR, and Berkeley NLP. I build AI systems that improve through experience, by optimizing their prompts, agent harnesses, and model weights.

My background in programming-language tools led to Monitor-Guided Decoding (MGD) at Microsoft Research. MGD built a static-analysis harness around an LLM to guide code generation, letting a smaller model outperform a larger one on code generation. This motivated my work at Berkeley on evaluating LLM systems with harnesses, rather than LLMs alone. In LangProBe, we mapped their cost–quality frontier and showed that the principle holds across many domains: smaller models with well-tuned harnesses can outperform much larger LLMs at lower cost.

This motivated exploring automated harness optimization from rich textual feedback. GEPA optimizes prompts by reflecting on execution traces and textual feedback, outperforming GRPO with fewer rollouts in our experiments; optimize_anything extends this to the full agent harness; and gskill learns repository-specific skills for coding agents. This led to studying how learning in text space could also support model training. mmGRPO is an RL method for multi-component LLM systems, combining weight updates with prompt optimization. Fast-Slow Training interleaves fast prompt adaptation with slower RL updates, so prompts absorb task-specific lessons while weights need not change as much. This improves data efficiency and performance while preserving the capacity of the model to learn subsequent tasks. I am continuing to study how self-improving AI systems should learn in the harness and in the model.