Carlos E. Jimenez

Jimenez joined Anthropic's research organization as a Member of Technical Staff in August 2025, according to his current CV. He completed a computer science PhD at Princeton in May 2026, advised by Karthik Narasimhan. Before Princeton, he studied at De Anza College and earned a computer science bachelor's degree at the University of Utah, where Ellen Riloff supervised his thesis. His stated research interests include autonomous systems, software engineering and collaboration between people and AI.[1][2]

Evaluating fixes with SWE-bench

Jimenez and John Yang share lead-author credit on SWE-bench. The original benchmark pairs repository issues with the code changes that resolved them. A system receives a codebase and an issue description, then must produce a fix that can be tested. This requires reasoning across an existing project, rather than answering a self-contained programming exercise. The paper gives readers a foundation for understanding what a coding-agent score measures and why a result on one task collection does not describe every kind of software work.[2][3]

Designing tools for SWE-agent

SWE-agent treats the language model as a user that needs a suitable computer interface. Its tools help the model locate files, inspect relevant code, make changes and execute tests. Jimenez co-led the paper, which examines how interface choices affect performance and behavior. The practical lesson is that improving an agent involves the tools and feedback around the model as well as the model itself. Its reported results describe the tested systems and benchmarks, not a guarantee that an agent will safely resolve an arbitrary production issue.[2][4]

Extending evaluation to visual software

SWE-bench Multimodal examines software issues that include images, such as defects in interfaces, diagrams and visualizations. Jimenez and his collaborators assembled JavaScript tasks where an image appears in the problem or tests. Systems successful on the original Python benchmark struggled on this collection. The work makes a useful distinction for teams choosing an agent: fixing backend code and understanding a visual user-interface failure exercise different capabilities.[5]

Creating training tasks with SWE-smith

Jimenez also co-authored SWE-smith, which generates software-repair training tasks. The system sets up an execution environment for a Python repository and creates changes that cause existing tests to fail. Those failures provide tasks for an agent to repair, making it possible to produce training examples beyond the limited supply of collected issue histories. The team released its procedure, tasks, trajectories and models so other researchers could study and extend the approach.[6]