Lakshya A. Agrawal is the creator of GEPA and a computer science PhD student at UC Berkeley, advised by Matei Zaharia and Dan Klein. His research examines how AI systems can use repository knowledge, execution traces, and diagnostic feedback to improve the code they generate and the instructions that guide them.
From language tools to repository-aware code generation
Agrawal studied computer science and applied mathematics at IIIT-Delhi. His early projects involved making programming languages work together: during Google Summer of Code in 2019, he developed Pytranslate, a Common Lisp transpiler that converts Maxima symbolic-computation code into Python. At EPFL, he built a GraalVM-based runtime for the CAL dataflow language, together with editor and debugger support. His career history also includes a software-development internship in Microsoft’s Azure Compute organization.
From 2022 to 2024, Agrawal worked as an AI4Code Research Fellow at Microsoft Research, investigating the quality and correctness of language-model-generated code. A central problem was repository-aware code generation: a model may produce plausible code while inventing a method or misunderstanding an API defined elsewhere in the project. Fluent generation alone does not supply that missing repository knowledge.
Agrawal is the first author of monitor-guided decoding, which brings static analysis into the generation process. A monitor constrains the model’s next choices using information about the repository, helping it select valid symbols, supply the correct number of arguments, and respect rules governing the order of API calls. The team constructed PragmaticCode, a Java method-completion dataset, to evaluate the approach against buildable repositories. In those experiments, a smaller model equipped with the monitor could outperform a much larger model on compilation and identifier-matching measures.
He developed multilspy to make language-server tools easier to reuse. The Python library handles server setup, configuration, and protocol communication while exposing operations such as finding definitions, locating references, and requesting type-aware completions. It gives AI coding applications access to repository knowledge that developers also use to navigate unfamiliar codebases.
Improving the system around the model
Agrawal began his Berkeley PhD in 2024. His research expanded from guiding individual code completions to evaluating and optimizing systems containing multiple model calls and tools. He co-authored LangProBe, a benchmark examining more than 2,000 combinations of tasks, program architectures, optimizers, and models. The work investigates cost–quality tradeoffs: optimized language programs can improve on direct model calls, but choosing how to compose them still requires empirical evaluation.
His contributions span several ways of adapting these systems:
Training modular agents: With Noah Ziems, Dilara Soylu, and Omar Khattab, Agrawal jointly led the development of dspy.GRPO. Its server–client design separates model-training updates from the execution of multi-stage agents, allowing reinforcement learning to accommodate programs with several specialized model calls and tool interactions.
Reflective prompt evolution: Agrawal led the development of GEPA, a collaborative prompt-optimization method that learns from execution traces. A model inspects reasoning, tool calls, tool outputs, and errors to diagnose failures and propose revised prompts. GEPA tests those revisions and maintains a Pareto pool of candidates that perform well on different training examples. Keeping complementary strategies in the search helps it avoid local optima and combine useful lessons, without requiring updates to the task model’s weights.
Optimization beyond prompts: In the collaborative optimize_anything research, Agrawal and his co-authors extend reflective optimization to text artifacts such as code, agent architectures, and scheduling algorithms. An evaluator supplies a score alongside diagnostic feedback to guide the next change. Their experiments include discovering an agent architecture that increased Gemini Flash’s ARC-AGI accuracy from 32.5% to 89.5% under the study’s conditions. Tests across three domains found that explanatory feedback produced faster convergence and higher final scores than scores alone.
Across these projects, Agrawal investigates how to make feedback useful for the next attempt. Static analysis supplies constraints while code is being generated; execution traces expose failures after an agent runs; evaluators guide revisions to the surrounding program. His work connects those signals to mechanisms for improving AI systems, from individual completions to entire agent architectures.
Lakshya A. Agrawal explains how full execution traces can teach an optimizer more than a reward alone—and how the same search can improve prompts, agent programs, repository skills and evaluators.
Reflection uses the diagnostic content of an attempt—intermediate outputs, tool responses and errors—to propose a behavioral change in text.
GEPA’s Pareto pool preserves candidates with strengths on individual examples, allowing search to continue along alternatives when the highest-scoring prompt stops improving.
Optimize Anything can search over prompts, programs and policies when they can be represented as text and scored. The evaluator should return actionable feedback alongside its score.
Learned repository skills can carry practical knowledge between attempts and models, reducing repeated exploration as well as improving task completion.
For subjective tasks, detailed human annotations can first train an LLM judge prompt; that judge can then guide agent optimization in a continuing production feedback cycle.
▶ Watch ↗