Scalable Intelligence through Generalist Multi-Agent systems. We study how many general-purpose agents can reason, plan, and build together — achieving what no single model can alone.
Human progress has always been collective. We believe artificial intelligence will follow the same path — through societies of capable, generalist agents that divide labor, verify one another, and compound their knowledge.
We characterize how performance scales not only with model size, but with the number, diversity, and organization of collaborating agents.
Rather than narrow specialists, we build broadly capable agents that can take on new roles, learn new tools, and transfer skills across tasks.
We design the protocols, incentives, and structures that let agents communicate, negotiate, critique, and converge on reliable outcomes.
Our work spans theory, systems, and evaluation — from the mathematics of collective reasoning to the infrastructure that runs thousands of agents in parallel.
Learned and structured protocols that let agents share intent, context, and partial results efficiently.
→How debate, decomposition, and peer review among agents yield more accurate and robust conclusions.
→Empirical and theoretical laws relating agent count, heterogeneity, and topology to system capability.
→Memory, planning, and self-correction for agent teams that work over hours, days, and complex projects.
→Methods to keep multi-agent systems aligned, auditable, and controllable as they grow more capable.
→Start with broadly capable foundation agents equipped with tools, memory, and reasoning.
Assign dynamic roles — planner, executor, critic, verifier — that adapt to the task.
Coordinate through shared workspaces and protocols that scale from a few agents to thousands.
Measure, verify, and amplify the capabilities that arise from the collective.
A single objective becomes a coordinated effort. Each agent contributes its strengths, every result is cross-checked, and the system converges on an answer it can stand behind.
We're looking for researchers, engineers, and collaborators who want to push the frontier of multi-agent AI.