How can multi-agent AI systems learn from their failures without retraining?

In May 2026, researchers from Oregon State University, Johnson & Johnson and other institutions published EVOCHAMBER, a framework in which a pool of AI agents keeps learning from its failures while the language model itself stays unchanged. After a failed or disputed task, the agents reflect together and pass lessons from strong to weak agents. With the open model Qwen3-8B, that raised accuracy on hard math problems to 63.9%, against 48.4% for the best comparable method. The price is about 3.6 times the compute of a single agent.

Why do agent teams forget what they learn?

A multi-agent system lets several instances of a language model work on a task together, each with its own role. According to the EVOCHAMBER paper, frameworks such as AutoGen, MetaGPT and CAMEL assign roles, but the agents' knowledge does not grow as new tasks come in. Memory methods such as Reflexion and ExpeL improve one agent at a time. MemCollab shares a single memory with every agent, which erases the differences that make a team useful. Methods that train agents together with reinforcement learning need weight updates and are frozen once deployed.

The authors of Solvita, a related system for programming contests, make the same point: multi-agent frameworks stay stateless and throw away the problem-solving and debugging experience of earlier tasks.

How does EVOCHAMBER evolve a team?

EVOCHAMBER changes prompts, memories, teams and the roster of agents, never the model weights. The authors call this test-time co-evolution and run it at three levels:

• Individual: after each task, every agent writes a lesson for that type of task (a "niche") and a general insight, and retrieves the most similar ones into its prompt next time. It also tracks a competence score per niche.
• Team: for each task, a selector picks three agents. The anchor is the best agent on that niche and also leads. The complement adds skills the anchor lacks, and the scout is an agent with little experience on the niche, which keeps the pool diverse. The leader chooses how the team works: voting, debate, generator and critic, or splitting the task.
• Population: after failures, CoDream moves knowledge between agents. Every few tasks, lifecycle operators copy strong agents with a more specialized persona, merge near-duplicates, remove agents that lag behind and create new agents for task types nobody covers.

In the experiments, the pool starts with 20 identical agents and teams of three.

What happens in a CoDream session?

CoDream (collaborative dreaming) starts when a team fails: the average score drops below a threshold, or the members disagree. The three members then go through five phases:

1. Reflect: each member diagnoses its own attempt.
2. Contrast: failing members are paired with successful ones to find what the successful approach did differently.
3. Imagine: those differences become candidate strategies, tagged with the niches they might apply to.
4. Debate: the members criticize each other's proposals and drop the weak ones.
5. Crystallize: the surviving proposals become structured insights, each with a scope: one task, a subdomain or all domains.

Each insight goes only to agents whose competence on that niche is below the pool median. Strong agents produce knowledge and weak agents receive it, so the specialists stay specialized. Sending every insight to every agent, as MemCollab does, dilutes that specialization.

What did the experiments show?

The authors tested EVOCHAMBER on three streams of tasks with Qwen3-8B:

• Hard math (382 problems from MATH and AIME 2022 to 2025): 63.9%, against 48.4% for MemCollab (32% better, relatively) and 29.8% for a single agent.
• Code (MBPP+ and CodeContests): 75.7% overall. On the hardest part, CodeContests, it reaches 35.2% against 6.8% for a single agent.
• Six domains in blocks of 100 tasks, from GSM8K to DROP: 87.1%, against 84.0% for the next best method.

Removing CoDream costs the most: on the six-domain stream accuracy drops from 87.1% to 76.3%. Shuffling the order of the math tasks did not hurt (65.5% and 66.2%). With GPT-4.1-mini as the model, the gain over a single agent is smaller: 10.7 points on hard math (34.1 with Qwen3-8B) and 1.9 points on the six domains. From 20 identical agents, four to five stable specialists emerge in each of three runs, but which agent becomes the specialist changes with the random seed.

Where else does the idea appear?

Solvita (14 May 2026) applies it to competitive programming. Four agents, a Planner, Solver, Oracle and Hacker, each get a trainable knowledge network shaped as a graph. Pass or fail verdicts and the weaknesses the Hacker finds update those networks through reinforcement learning, while the language model stays unchanged. The authors report that Solvita beats existing multi-agent pipelines on CodeContests, APPS, AetherCode and live Codeforces rounds, and nearly doubles the accuracy of single-pass baselines.

A much smaller experiment runs on modest hardware without a GPU. flux-genotype, which one developer presented on Reddit on 18 May 2026 as an alpha, runs local models through Ollama on a CPU. A judge model scores each answer, and when the score drops below about 0.618, a third model rewrites the system's architecture file. In one run it replaced the judge itself, Llama 3.2, with DeepSeek-Coder 6.7B.

What are the limits?

EVOCHAMBER uses about 3.6 times the inference of a single agent, which the authors call possibly prohibitive where latency matters, although it beats five-way majority voting on 72% of that method's token budget. It was tested with two model families, on streams of up to 600 tasks, and its lifecycle operators use fixed thresholds. The authors name longer task streams, stronger models and insights that go out of date as open questions. The code is on GitHub under the MIT license.