The starting problem is concrete: an AI agent can write a patch that looks plausible and still violates the physical assumptions inside a model. For Earth system modeling, a passing build is not enough. A code edit can change water balance, energy closure, or the meaning of a parameterization.
That problem runs through Koutian Wu's current work at UT Austin. His primary expertise sits at the intersection of geophysics modeling and artificial intelligence. He studies how machine learning and agent systems can help with Earth system models, land surface models, and space physics workflows while keeping the physics visible to human review.
ESM-bench is the clearest expression of that concern. The benchmark asks whether AI agents can modify Earth System Model code while respecting both software structure and physical logic. In Wu's April 2026 post, the benchmark is described as 576 tasks across physics-based bug fixes, process representation modifications, parameterization scheme selection, and parameter optimization. The same post reports a code-level hint run where Claude Opus 4.7 reached 0.312 structural F1 under an assisted v4 Noah-MP setting, while exact match stayed at 0 percent.
The point is diagnostic. A model may find a plausible file, write a syntactically clean patch, and still miss the physical mechanism. Wu's evaluation design separates localization, patch generation, evidence use, and physics-aware review so the failure mode can be seen instead of hidden behind one score.
Noah-Agent extends the same idea into climate-model workflows. It frames automated parameterization and validation for Noah-MP as a multi-expert agent process. The target is not a chatbot around a model manual. It is a workflow that can inspect Fortran code, reason about configurations, connect observations to model output, and leave a trail that a domain scientist can audit.
Wu's agent work also comes from daily use. He has used coding agents across research and software work for nearly 300 days. His typical workflow gives Codex and Google Antigravity the planning role, uses Claude Code for fast implementation, then brings Codex CLI back as an independent reviewer before rebasing and merging. The workflow is productive, but it also exposes the weak points: fabricated citations, wrong model identifiers, bad PR scope, and long exploration when a small verified edit would have been enough.
That experience feeds into ResearchSkills.ai, where Wu is a co-lead. The project collects researchers' decision trajectories and turns them into reusable agent skills. His recent work, Commit Log as Skill Mine, treats scientific commit histories as data for extracting reusable know-how from research workflows. In this view, an agent skill is not a prompt trick. It is a small, testable piece of procedural knowledge.
The evaluation thread also appears outside climate modeling. At PineAI, Wu built a combined LLM-as-a-judge and rule-based evaluation pipeline for an agent knowledge base. The system processed more than 1,000 call sessions, removed more than 2,000 noise or PII entries, extracted more than 3,000 Q&A pairs, and evaluated outputs for relevance, accuracy, completeness, consistency, and clarity. That work made the failure modes of judge models a practical engineering issue rather than an abstract concern.
His scientific ML work follows the same habit of asking where a model is allowed to be uncertain. In a Texas water-quality project, he paired Sentinel-2 imagery with NERRS in-situ observations to retrieve turbidity and chlorophyll-a for coastal bays. The pipeline benchmarks 12 regressors, builds spatial prediction maps with selected GBDT and Ridge models, and tracks transferability because the strict matchup dataset is still small.
Before this AI-agent work, Wu published in JGR: Space Physics on diurnal and seasonal variations of meteor speeds from Mengcheng Meteor Radar observations. That project used a multi-stage signal-processing pipeline to filter interference from raw meteor radar data. It is a useful anchor for his later AI work: the point is not to make models fluent, but to make them reliable in scientific settings where bad confidence can look like a result.
On earth-space-ai.org, that stance becomes operational. Wu serves on the executive committee, working on releases, issue triage, docs, partner outreach, funder relations, and research-integrity review. The same pattern appears across the site: make expert procedural knowledge loadable by agents, readable by humans, and reviewable by the community.