Intelligence grows
when agents learn together.

We build collaborative AI agents that turn shared experience into greater capability—tackling frontier problems where the best path has yet to be discovered.

Local sharing · memory / evidence / skillsAgent migration between subnets →
GENERATION 04
AGENTS16SUBNETS04INTELLIGENCE1.00×
CURRENT NETWORK STATEFour subnets online. Local sharing is active.
Agents share locallyIntelligence compoundsAgents migrate between subnets
THE COMPOUNDING LOOP

More agents generate more experience.
Shared experience compounds intelligence.

Multi-agent systems widen the search and multiply what can be learned. Continual learning turns every outcome into capability, then returns it to the agents.

01 / MULTI-AGENT

Expand intelligence.

Specialized agents explore in parallel, share discoveries, and collect experience at a scale no single agent can.

Parallel searchShared evidence
02 / CONTINUAL LEARNING

Learn from experience.

Every result becomes memory, skills, and better strategies—returned to the network for the next generation.

Retained memoryImproved strategy
PROOF POINT · KERNEL ENGINEERING

On Anthropic's public GPU kernel benchmark.

19%

faster.

Four agents. One night. No humans.

WALL-CLOCK PERFORMANCECORAL pushes beyond the previous best-known score.
OpenEvolveCORAL · 1 agentCORAL · 4 agents
ModelBase modelEvals ↓Best cycles ↓Speedup ↑Imp. rate ↑
Best KnownAnthropic—1363108.4×—
OpenEvolveOpus 4.6363274053.9×3% (12/363)
CORAL · 1 agentOpus 4.6561350109.4×43% (24/56)
CORAL · 4 agentsOpus 4.65961103133.9×9% (54/596)
Anthropic kernel engineering task · best known normalized to 1.0
OPEN-SOURCE INFRASTRUCTURE · APACHE 2.0

CORAL

Autonomous AI agent organizations that run experiments, share knowledge, evaluate results, and continuously improve solutions.

Claude CodeCodexCursorOpenCodeKiro
DocumentationView on GitHub
RESEARCH FOUNDATION

Work that
compounds.

Our research spans agent memory, economic coordination, reasoning, and multi-agent evolution. Each result becomes infrastructure for what follows.

01

MEM1

Long-horizon agent memory.

NeurIPS MTI-LLM Best Paper · ICLR
02

Economy of Minds

Emergent intelligence through economic interaction.

Multi-agent intelligence
03

CORAL

Multi-agent evolution.

8 new state-of-the-art results
COLLABORATORS ACROSS SECTORS

Intelligence compounds
Across frontiers.

01 / COLLABORATOR FIELDHome EnergySystems & control
02 / COLLABORATOR FIELDPharmaResearch & discovery
03 / COLLABORATOR FIELDKernel EngineeringPerformance & optimization
04 / COLLABORATOR FIELDRecommender SystemsModels & personalization
05 / COLLABORATOR FIELDLogisticsRouting & operations

We work alongside domain experts where progress depends on deep specialization—bringing compounding intelligence to science, energy, and engineering.

Start a conversation