Cheaper
Low-cost models carry hard work
When each agent owns a small part, a low-cost model can do it well. A graph of open models beat the strongest frontier model on hard codebase questions, at half the cost per correct answer.
Coral AI Labs / Organised intelligence
AI gets more capable when it's organised.
Coral AI Labs researches and builds the infrastructure for organising AI agents on any model, so cheap intelligence can finish hard work correctly.
Ahead of the labs' strongest single models on hard software and general tasks, and of MIT's research system on algorithm search.
Organisation, at scaleIllustrative agent graph
216agents
Many responsibilities.
One connected piece of work.
Many focused agents investigate in parallel. Each holds one part of the work.
ActiveSleeping
6 illustrative responsibilitiesResearch → Infrastructure → ApplicationsThe bigger picture
02 / The shift
AI got cheap. Getting it right didn't.
For five years, AI has improved one way: bigger models, longer context, more thinking per answer. It has worked. The price of a fixed level of AI performance now falls about thirteen-fold a year.
But real work is bigger than one model's attention. A codebase, a research programme or a company's operations is connected, and it changes as you work on it. A single agent holds one context and follows one line of thought. Give it an hour-long task and it fills up with dead ends, then answers confidently and wrong.
The cost lands on people. When firms adopted AI coding agents, they submitted 36% more code but resolved only 14% more issues, and code waited 56% longer for someone to approve it. Output went up. Finished work barely moved.
13×
cheaper each year for the same AI performance
+56%
longer wait for code approval after firms adopted AI agents
2 h 23 m
for an engineer to find and fix one wrong answer about unfamiliar code
03 / Our thesis
Intelligence has a horizontal scaling axis.
Each lab makes AI more capable by making its own models bigger. We study the other direction: how many agents, often on different models and built by different teams, work one problem together.
The gain comes from how agents are organised, not how many there are. Each agent owns a responsibility. Agents pass evidence to whoever needs it, challenge each other's conclusions and revise the plan as they learn. Done badly, extra agents duplicate work and leave a person to sort it out. Done well, they finish work no single model can.
We measured the difference. Given the same extra compute, independent attempts at a task reached 37.9%. A coordinated graph of agents reached 62.1%.
Bigger models raise the ceiling. Organisation decides how much of it you reach.
SourcesAgentRadio
04 / Why it matters
What changes when AI is organised.
More reliable
Agents check each other
Agents challenge conclusions before they reach a person, so fewer confident, wrong answers land on someone's desk.
Safer
Narrow agents, clear limits
Each agent has one responsibility, its own tools and its own access, revoked when the work ends. An independent review rated Coral more secure than Google's A2A and the Linux Foundation's ACP. More capable systems should not mean less safe ones.
Yours
Capability you own
Coral runs on any model and inside your own infrastructure. What your AI can do stops depending on one lab's next release, prices or rules.
Every lab gains too. Organised agents put more of each lab's models to work, on tasks that now succeed.
05 / What we've shown
Proved in three different kinds of work.
General tasks
+8.4 ptsAhead of the strongest open-source workflow system on GAIA, with the same roles and models
See the resultsSoftware
+25 ptsCoral Code answered 76.6% of hard codebase questions fully. The labs' best single model answered 51.6%.
Research
275 vs 596Evaluations Coral Discovery needed to reach MIT's best score on a published kernel problem, with the same model and team size
See the resultsEach result came from organisation, not a bigger model. The same approach works on whichever model leads next quarter.
06 / Research
An open research lab.
We publish what we learn, with co-authors at universities in the UK, Europe and the US, and we build on an open core anyone can run.
Coral Discovery: The Division of Cognitive Labor in Autonomous Research
Preprint, with Trento, Harvard, Cambridge, King's College London and Hull
A shared direction board lets many agents search a problem without converging too early
AgentRadio
Under review, AAAI-27
Negotiated plans and passive awareness took one model from 32.3% to 62.1% on codebase questions
Beyond Rule-Based Workflows
Preprint
Information routed as evidence arrives beats fixed multi-agent workflows
Anemoi
Peer reviewed, NeurIPS 2025
Agents that talk to each other directly beat a central planner relaying everything
Coral co-authored Permission Manifests for Web Agents with researchers from Oxford, Stanford, Johns Hopkins, Microsoft, UT Austin and GovAI, among others.
What others found (independent, no Coral authors)
Samsung Research Montreal and McGill reviewed 18 agent protocols and described Coral as moving multi-agent work toward "verifiable meaning, trust, and incentive alignment". Beyond Message Passing
Ariel University rated Coral more secure than Google's A2A and the Linux Foundation's ACP. Security analysis
07 / What we build
One engine. Applications that prove it.
Research becomes infrastructure, and infrastructure becomes products people use every day.
Coral Cloud
Build an agent team on a canvas, then watch it talk
Coral Cloud runs Coral Server for you. Pick agents from the marketplace, draw the groups that decide who can talk to whom, and start the session. Every message between agents shows up live, so you can see who asked what and step in when a thread drifts. Save a graph as a template and start the next one from it.
- The graph is the wiring.
- An agent only talks to agents in its own group, so the graph you draw is the conversation you get.
- Spend has a ceiling.
- Sessions draw on a prepaid balance with a limit you set. Agents reach models through Coral Cloud, so no provider key sits inside an agent.
Open Coral Cloud Free tier available
Field recording / Coral ConsoleProduct footage
- 01Add agents
- 02Group into a graph
- 03Load a template
- 04Watch the threads
Coral Server
Teams building AI systems
The engine that runs graphs of agents on any model, inside your own infrastructure
08 / Where this goes
From agents that work together to systems that remember.
What we're working on
Memory that stays with the work.
Today an agent forgets when its session ends. We are testing graphs that keep what they learn, so every new task on the same work starts warm and costs less.
What we're working on
Small models, made for their part.
Small models are cheap to tune. We are combining and tuning them for the responsibility each agent holds, so reliability comes from organisation, not size.
What we're working on
Intelligence you can compose.
We want people to assemble, reshape and govern systems of many models as easily as they run software in the cloud today.
Bigger models made individual AI more capable. Organised AI can make it abundant, dependable and safe across every kind of work.
Build with us.
We work with research teams, companies with hard problems, and investors who see where AI is going.