Stop guessing your agent architecture. Evolve it.
Supremum is an evolutionary discovery engine for multi-agent orchestration. We find the optimal topology for your specific task through rigorous search and multi-perspective evaluation.
Every framework ships one default topology. That's the bottleneck.
Flat hierarchy. Broadcast communication. Consensus decision-making. It's the "no organizational design" default — and it's leaving performance on the table for every multi-agent deployment in production.
Default Agent Architecture
- close Every agent sees everything (broadcast noise)
- close Outputs concatenated with equal weight
- close Same topology for every task type
- close Redundant token spend across agents
After Organizational Search
We treat agent team design as a search problem over the space of all possible organizational configurations, evolving structures that outperform human intuition and framework defaults.
With Supremum
-
checkInformation flows match task requirements
-
checkDecision protocol shapes output assembly
-
checkLoad-bearing structures identified via ablation
-
checkValidated against multiple baselines
Three layers of rigor most agent tools don't attempt.
Evolutionary Search
We evaluate thousands of organizational configurations — hierarchy, communication topology, decision protocols, role composition — to find the global maximum for your task.
Adversarial Evaluation
Outputs are scored by a multi-model, 5-seat evaluation council designed to eliminate single-LLM judge bias, position effects, and hallucinated quality signals.
Causal Ablation
We don't just give you a structure. We ablate every architectural gene to show you exactly which decisions are load-bearing and which are decorative.
Not a recommendation. A sensitivity analysis.
Every gene in the winning genome is independently tested. You know precisely which architectural decisions matter for your specific task.
| Gene | Winner | Fitness Impact | Verdict |
|---|---|---|---|
| Hierarchy | Hub-Spoke | −0.12 | Load-bearing |
| Communication | Chain | −0.08 | Load-bearing |
| Decision-Making | Leader-Decides | −0.02 | Decorative |
| Work Distribution | Specialized | −0.06 | Load-bearing |
Pre-registered. Falsifiable. Honest.
Supremum is backed by a formal experimental protocol with explicit refutation conditions. Every claim ships with the conditions under which we'd retract it.
Organizational Architecture Search: Evolutionary Discovery of Optimal Multi-Agent Team Structures
Robinson, N. & Magnus†. 2026. Pre-registered protocol with 5 falsifiable claims, 15 tasks across 5 cognitive categories, 63 evolutionary runs, human-calibrated evaluation, three-tier baseline comparison.
† Magnus is a persistent cognitive agent. Co-developed the protocol, evaluation methodology, and experimental design.
In Progress — 2026Your topology is silently determining output quality — and most teams never test it.
No single topology dominates. The optimal structure depends on the task — crisis response and code review want different architectures.
Evolutionary search finds structures that outperform both naive defaults and expert-designed configurations.
Most architectural decisions don't matter. A few are load-bearing. We tell you which is which.
We're onboarding design partners.
If you're deploying multi-agent systems in production and hitting architectural performance bottlenecks — inconsistent output quality, unexplained cost variance, topology decisions made by intuition — we want to hear from you.
What design partners get: We run the Topology Engine on your production task and deliver a sensitivity analysis showing which architectural decisions are load-bearing, which are decorative, and the specific genome that outperforms your current setup.
Model-agnostic. Works with any LLM provider. Runs cost $15–25 per task.
Request Access