Thesis.

Two research agents can spend a week on the same problem and never find out that the other one solved it. The useful part of the first agent’s work might be a failed approach, a small change to a protocol, or a result that only holds under certain conditions. If that knowledge stays in its original workspace, the second agent has to pay to learn it again.

Research already has a way to carry discoveries across teams. A paper makes a result public, and citations and follow-up experiments help others decide how much to rely on it. Agents need a way to do the same thing at the level of individual findings, with enough context for the next agent to use them. That is what we mean by a hive mind for agents.

How agents connect through Spock.

Trust comes before exchange.

Imagine an agent finds a claim that a method works on a new dataset. To reuse it, the agent needs to know how the dataset was collected, which examples were held out, and whether the reported result depends on a setting the authors left out. The system should be able to find the relevant evidence and ask the publishing agent about the gaps. A finding becomes more useful when its evidence and limits travel with it.

The team that did the work should decide what gets shared. An agent can propose a write-up of a useful finding, and the team can review it before anything leaves its workspace. Over time, we want agents to find verified answers from other teams and give the people who produced them a way to be paid.

We’re starting with scientific search.

Our first step is to help agents find evidence in the scientific work that is already public. A search result tells an agent which paper to open. SEL, the model we’ve released, helps it find the passage or region worth inspecting. That makes it cheaper to check several sources when a claim depends on more than one of them.

From there, the next problem is verification. We want agents to be able to ask how a result was obtained and decide whether it applies to their own work. Once they can do that, a result from one workspace can become a starting point for another.