When AI agents work together on scientific ideas, can you change their teamwork rules one at a time and measure each effect?

This explores whether the rules for how AI agents collaborate on scientific hypotheses can be isolated and treated like experimental variables, so you can change one and see what it does. The clearest yes in the collection comes from HypoEvolve. It treats hypothesis development as a genetic algorithm: a population of ideas goes through selection, recombination and mutation over generations. That framing separates what each agent does scientifically from the rules that decide how they coordinate. Because the rules are written down instead of buried in prompts, each one can be named, swapped out and compared. On drug repurposing tasks it beat six baselines How do collaboration rules shape hypothesis quality?. The broader point is that the evolutionary framing is what makes the rules testable, not just the ingredient that makes the system work.

Compare Google's Co-Scientist, which also has a generate-debate-evolve loop, organized as a tournament where hypotheses compete and earn Elo ratings. Its builders report that hypothesis quality rises with more compute Does more thinking time improve AI-generated research hypotheses?. That result is about turning up the budget, though, not about changing the tournament's rules. Robin, a lab-in-the-loop system, adds a different kind of collaboration rule: it runs 10 independent attempts and takes a consensus across them, with human experimenters closing the loop Can multi-agent systems guide wet-lab discovery through iterative cycles?. Each system builds in a different choice about how ideas compete or combine. Explicit rules would let you test those choices against each other instead of judging the systems as wholes.

The opposite approach shows why explicitness matters. Reasoning models given a shared working memory (a concurrent KV cache) start coordinating on their own. They make plans, notice duplicated work and divide labor, with no rules and no training Can multiple LLMs coordinate without explicit collaboration rules?. That is impressive, but there is nothing to vary: when the coordination is emergent, you can't switch off one behavior to see what it contributed. A related idea appears in weight space, where swarms of models search for new expert combinations using particle-swarm rules Can language models discover new expertise through collaborative weight search?. Population-search framings like this seem to be a recurring way of making collaboration inspectable.

The less obvious reason this matters is that leaving collaboration to defaults often goes badly. Frontier models that solve problems well alone do worse when they collaborate, agreeing with each other more than 90% of the time whether or not they're right Why do language models fail at collaborative reasoning?. LLM groups also settle on an answer sooner and bring up less unique information than human groups Do language model groups mimic human group reasoning patterns?. Team makeup is another variable that interacts with the rules: diverse teams help only when members have real domain expertise, and diverse non-experts do worse than one competent agent Does cognitive diversity alone improve multi-agent ideation quality?. So collaboration rules can be tested independently, but their effects probably depend on who the agents are. The collection doesn't yet have a study that varies rules and expertise together. That gap looks like the obvious next experiment.