What if AI agents could evolve together?
Most multi-agent AI systems work like a team. One agent plans. Another writes code. Another tests it. Another reviews the result. Each agent has a predefined role.
It’s a sensible design and it borrows straight from how human teams work. Assign roles, hand off between them, done. But notice what’s fixed in that picture. The roles don’t change. The planner plans forever, the tester tests forever, and none of them gets any better at the actual problem over time. The team executes a workflow. It doesn’t grow. Run it a hundred times and the hundredth run is no smarter than the first, because nothing carries forward except the output.
A different approach
But a new framework called CORAL takes a different approach. Instead of simply collaborating, multiple AI agents independently explore different solutions, experiment, learn from their results, share discoveries through long-term memory and continuously improve over time.
The shift is from division of labour to shared exploration. Rather than each agent owning one step of a pipeline, multiple agents go off and attack the problem independently, each trying its own approach. They experiment. They see what their attempt produced. And here’s the part that changes everything, they write what they learned into a long-term memory the others can read. So a discovery made by one agent doesn’t die with that agent’s run. It becomes something the whole group can build on. The memory is the connective tissue. It turns a set of parallel attempts into a group that accumulates knowledge.
They don’t just work together. They co-evolve.
That’s the distinction worth holding onto. Working together is dividing a task. Co-evolving is the agents getting collectively better at solving it, run after run, because each one’s progress feeds the others.
Keeping the good, dropping the bad
If one agent gets stuck, a heartbeat mechanism can detect it and redirect or restart it. Good ideas are preserved. Bad ideas are discarded. Other agents build on successful discoveries.
This is the selection pressure that makes “evolve” more than a metaphor. An agent that stalls is a dead end burning compute, so the heartbeat mechanism watches for exactly that, notices when an agent has gone quiet or unproductive, and redirects or restarts it instead of letting it hang. Meanwhile the ideas get filtered. What worked is kept in memory and becomes a foundation. What failed is thrown out so nobody wastes another cycle rediscovering that it doesn’t work. And the successes compound, because the next agent starts from a discovery rather than from scratch. Keep the good, cut the bad, build on what survives. That’s evolution running inside a group of agents.
What it actually did
The researchers tested CORAL across mathematics, algorithms and system optimisation. It achieved the best results in 10 tasks and produced 3-10x more improvements than traditional evolutionary methods while using fewer evaluations.
The numbers are the part that makes this more than a nice idea. Best results across 10 tasks spanning three different domains, so it’s not tuned to one narrow problem. And 3 to 10 times more improvement than traditional evolutionary methods, which is the fair comparison, since those methods also keep good solutions and discard bad ones. The difference is the shared memory and the co-evolution on top. What seals it is “while using fewer evaluations.” More improvement for less compute. Usually you buy better results by spending more. CORAL got better results while spending less, which is what tells you the structure itself is doing real work.