The idea¶
Most writing about AI in science covers large systems, such as autonomous "co-scientists" and agents that write papers. That is doing research. This site covers running research as well: choosing and planning projects, keeping track of data, code and notes, running a lab's routines, and the reporting, budgets and paperwork around science. It records which tasks scientists give to agents, how they check the results, and what went wrong.
One rule¶
The freedom an agent gets should depend less on how hard a task is and more on how cheaply its output can be checked. Delegate by verifiability explains the rule, and each recipe states how its output is checked.
The Comment behind the site¶
The recipes accompany a Comment, submitted for publication:
Theis, F. J. & Helmholtz Munich Computational Health Center. Delegate what you can check: ten lessons on AI agents in research.
Recipe N on this site follows point N of the Comment, with the steps, the check and the known problems written out in more detail than a Comment allows. A link to the Comment follows on publication.
Where the first recipes come from¶
The Helmholtz Munich Computational Health Center brings together about 45 research groups in computational biology, machine learning and clinical genomics. In September 2026, about 30 of its group leaders spent a retreat discussing what had worked with agents. A survey of all group leaders adds numbers: 36 of about 45 replied. The centre wants to use agents in all parts of its research work, from choosing projects to administration. It is not there yet, and some groups are further along than others.
Names and details are removed from all examples. The examples come from one centre with a computational focus, and are based on what people reported. People who already use agents a lot are overrepresented. Other groups can test each recipe, add their own examples, and report where a recipe does not work for them.
Skills as briefs¶
Most useful skills depend on the institute they were built at: its forms, its cluster, its rules. So the site publishes a brief for each skill instead of one institute's version. A reader gives the brief to their own agent, answers its questions about their institute, and gets a local version.
Relation to aiforscience.eu¶
aiforscience.eu is a living record of best practices for institutions using agentic AI in science, maintained by task forces and institutions. Its practices state what should be done. The recipes on this site show how to do it in a lab. Each recipe links to the practices it follows, and practices can link to recipes as examples. Both sites are maintained together and updated as the field develops.
Who it is for¶
- Scientists who use agents and want to use them for more tasks, safely.
- Group leaders who decide what their group should try next.
- Core facilities and institutes who decide which skills and systems to support.
What institutes can do now¶
Individual scientists start using agents without help. Much of what helped could not have been built by any single group. Institutes and funders can provide these now, without waiting for better models:
- Well-posed problems, each with an agreed metric and several datasets.
- A shared library of reviewed skills.
- Sandboxed computing.
- Documentation that agents can read.
- Peer learning.
- Capable tools for administrative staff.
- Agent interfaces on top of the existing finance, personnel, grant and travel systems, with read access where the rules allow and a person approving anything that is submitted.