Did the final artifact beat the strongest individual proposal?
Open source / private research lab
Measure theswarm.
SwarmSim is an open-source protocol lab for multi-model AI. It explores how independent agents gather, disagree, and converge, then measures whether the collective beats its strongest member.
—joined the private waitlist
The thesis
More models do not automatically mean a better answer.
AI councils can correct mistakes. They can also amplify them, suppress a correct minority, or spend five times more to produce the same result.
SwarmSim makes the interaction visible and compares every final answer against the strongest single model and a cost-matched baseline.
Did the models agree while relying on conflicting evidence?
Is the improvement worth the additional calls and latency?
A protocol, not a prompt
Design how intelligence moves.
Give each model a role, control what it can see, preserve useful disagreement, and decide when another call is worth making.
- 01 Bring models from different providers
- 02 Compare reusable collaboration protocols
- 03 Replay every critique and revision
- 04 Share honest, reproducible experiments
protocol: blind-peer-reviewbudget:max_cost: $2.00participants: - analyst # independent- critic # anonymous- verifier # evidence-firstflow:propose → review → verify↓synthesizeevaluate: - strongest_single- cost_matched_best_of_n- human_preference Protocol validEstimated 8–12 calls
Designed for consequential work
One lab. Many boards.
This private lab starts with research review, product decisions, and protocol evaluation on real tasks.
Research review
critic + verifier↗Stress-test a proposal before submission.
Product decisions
panel + synthesis↗Surface assumptions and preserve dissent.
AI evaluation
benchmark + replay↗Compare protocols on your own repeated tasks.
Questions, answered
Built carefully. Shared early.
Is SwarmSim a company?
Not currently. SwarmSim is an independent private project and research lab exploring measurable multi-model collaboration.
Is this another AI council?
Not quite. Councils give you a combined answer. SwarmSim is designed to show whether the collaboration improved the answer, where it failed, and what the improvement cost.
Which models will it support?
The first preview will focus on connecting models through OpenRouter, followed by direct provider and local-model support.
Who is the early cohort for?
Researchers, AI product builders, and people doing high-value knowledge work who want to test multi-model workflows on real tasks.
When will access open?
Small private cohorts will be invited as the evaluation harness and initial protocols are validated.
A better answer is not enough
Know why it is better.
Follow the private lab as it develops, and join the first small cohort of SwarmSim experiments.