for scientific discovery

We post-train AI agents for scientific discovery, and we study how to make that work better.

AI research

Post-training
for discovery

We train AI agents so they can take part in scientific discovery. Around each agent we build a harness: the tools, the memory, and the checks.

Then we study that process. How to make the training and the harness better. How an agent learns the data of a field. How several agents can collaborate on one scientific question.

Better training

We study how to post-train an agent, and how to make that training more effective.

Know the field

We study how an agent comes to know the data of a field, and how that knowledge should be trained in.

Work together

We study how several agents can collaborate on one scientific question.

What we do
in the lab

We train an agent, put a harness around it, and study what makes it better at scientific work.

The agent – Post-trained for scientific work, so it knows the data and the questions of a field.

The harness – Tools, memory, and checks around the model. We build the harness, and we study which designs make the agent more capable and more reliable.

The group – Several agents on one scientific question. We study how they should collaborate, and how to make that collaboration work.

We write the experiments up. The posts show the setups, the tests, and what we are learning.

A dandelion clock releasing moments of research into the past and the future

Posts

agentsscience

Swarms of agents spent weeks on the work. Human review started to fall behind.

Recent articles

View all

modelsdecisions

I put decision models in a house with mid-level LLMs. They took the scoreboard and the clock.

4Read more →

We start withenvironmental science

The first field these agents are trained for. We study the harness, how an agent learns the field's data, and how agents collaborate on a discovery.