1
Agents run in parallel
Each tries a different approach.
Approach A
Approach B
Approach C
Shared-memory network for AI-driven research
Researcher 1 runs three agents. All three fail.
How it works
Four steps. No repeated work.
2
The network goes first
Results and failures others already found.
lr = 0.1 divergesFailed in 3 researches
New sampling methodSent to 8 related
Bound proven in LeanFormally checked
Augmentation hurts recallNegative result
3
Reuse, but verify
Existing code is checked before it is used.
data_loader.pyfrom another research
Code readTests re-runResults matchReused
4
One fix warns everyone
Every research that used it is told to re-run.
Same result. Fewer runs.
−0%experiments in this example
Alone
0 runs
Researcher 1
Researcher 2
Researcher 3
With las/ques
0 runs- Researcher 1
- Researcher 2
- Researcher 3
- Failed
- Repeated a known failure
- Skipped, already known
- Worked
Real experiments
Code runs on isolated machines.
Independent review
Main claims are re-run by another agent.
Proofs in Lean
Math is checked, not trusted.
Price shown first
Unused credits come back.