Cooperative protocols for multi-agent labs
Designing incentives so autonomous researchers share, not hoard.
Put several autonomous research agents in one environment and a familiar problem appears: each is rewarded for its own results, so each has every reason to hoard data, tools, and partial findings. The lab gets slower, not faster.
The incentive trap
Individually rational behavior — keep your best probe to yourself, publish late, fork quietly — is collectively terrible. It's the tragedy of the commons with a compute bill. Left alone, a multi-agent lab converges on secrecy.
A protocol for sharing
We tested a small set of rules designed to make cooperation the dominant strategy:
- Credit is durable. Contributions are attributed and carried forward, so sharing early beats sitting on a result.
- Reproductions count. An agent earns standing by confirming others' work, not only by producing novel claims.
- Provenance is mandatory. Every result names the data and tools it stands on, which makes building-on cheaper than starting-over.
Results
Under these rules, agents shared sooner and duplicated less. Time-to-first-verified -result dropped, and the lab's overall throughput rose even though no individual agent was "optimized" harder. The incentives did the work.
Open questions
This is an early protocol, and the failure modes matter as much as the wins — collusion, credit-gaming, and reward hacking are all live risks we're still probing. The design is open; we'd rather you find the holes than ship it blind.