AI agents are already changing how climate researchers work

AI agents are helping climate researchers update scientific code, build forecasting systems and shorten the path from analysis to action. A Columbia Climate School event recap describes promising applications, alongside a useful warning: faster output makes verification more important, not less.

Watch Desk analysis

What happened

At a Columbia Engineering event in late September, researchers and industry experts discussed how agentic AI is being used in Earth science. Columbia Climate School’s recap describes systems that can write and run code, analyse data and iterate with limited human oversight.

Google Research scientist Ignacio Lopez Gomez said an agentic framework helped develop a California forecasting system that outperformed the operational forecasts then in use. He said the work took one to two months. Columbia professor Pierre Gentine described agents converting older Earth-system modelling code into Python that could use GPU acceleration. These are speakers’ accounts in the Columbia event recap, not independently evaluated results.

Why it matters

Climate models and forecasting systems depend on complex code that can take substantial effort to update. If agents make code migration, analysis and testing quicker, researchers may be able to try more approaches and put useful tools in reach of more people.

The recap also shows why speed is not the whole story. AWS architect Sameer Mohammed warned that code changes can introduce silent failures, including changes in how a program handles state. Researchers at the event argued that agents can also help write tests and verify results, but that still leaves people responsible for deciding whether the work is sound.

Our read

This is a better case for agentic AI than a demo of a machine producing code at speed: researchers are describing concrete work in climate modelling and forecasting. The important question is whether the gains survive careful checking and use beyond the examples discussed at the event. A scientific result needs more than code that runs; it needs code that does what researchers think it does.

What to watch

• Whether the California forecasting result is documented and evaluated beyond the event account.
• How researchers check agent-written or migrated code for silent failures.
• Whether these tools make climate research more accessible, or mainly increase the amount of output experts must review.

Discussion spark: Should climate researchers let AI agents make substantial changes to scientific code, provided the results pass rigorous testing, or should human review remain mandatory at every step?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.