Agents that learn from each other, within an episode and across episodes.

The loop

Execute, evaluate, evolve

Every agent runs the same three steps in every episode. Below, each step shows the call your system makes and what it actually produced in the recorded episode above.

memory = plugin.context("agent_c")
reply  = agent_c.solve(problem, memory)
plugin.observe("agent_c", round_messages)

result = check(final_answers, answer_key)
# task-native: accuracy, deal value, win/loss…

plugin.end_episode("agent_a", result)

Memory and knowledge

Two speeds of learning

Memory is updated during an episode from what the agent sees. When the next episode starts, working memory and facts marked as belonging to the last episode are cleared; past events and longer-lived facts stay, and a fact that is not seen again for a few episodes expires. Knowledge outlives the episode and is revised by each outcome. Here is Agent A's state across three consecutive problems of the recorded run.

Try it

Run the plug-in in your browser

The real evolvemas package runs here with Pyodide. Scripted agents need no key; switch to a live model to use your own.

Nothing yet.

Results

One protocol across cooperative, mixed-motive and adversarial settings

Use it

Three calls around the loop you already have

from evolvemas import EvolveMAS, AnthropicLLM

plugin = EvolveMAS("runs/memory", AnthropicLLM())
for a in agents:
    plugin.register(a.name, profile=a.profile)

for problem in problems:
    for a in agents: plugin.begin_episode(a.name)
    for rnd in range(2):
        replies = {a.name: a.solve(problem, plugin.context(a.name))
                   for a in agents}
        events = [{"speaker": name, "content": text}
                  for name, text in replies.items()]
        for a in agents:
            plugin.observe(a.name, events)
    result = check(replies, problem.answer)
    for a in agents: plugin.end_episode(a.name, result)

Paper

EvolveMAS: A Unified Self-Evolution Framework for Decentralized Multi-Agent Systems

Decentralized multi-agent systems let LLM agents interact without central control, across cooperative, mixed-motive and adversarial settings. EvolveMAS supports test-time adaptation within and across episodes with one protocol: a dedicated MemoryAgent keeps each agent's Structured Memory outside its reasoning context, and each episode's outcome is distilled into reusable partner profiles and interaction strategies. Across five benchmarks and six configurations it reaches the highest normalized average score among the compared methods, 77.2.

The plug-in on this site is a compact, host-agnostic implementation of the protocol. It is not the paper's experiment code.

BibTeX

@article{wu2026evolvemas,
  title   = {EvolveMAS: A Unified Self-Evolution Framework for Decentralized Multi-Agent Systems},
  author  = {Wu, Yalun and Ma, Qiwei and Xu, Bingbing and Zhang, Dan and Bai, Long and
             Wang, Boyang and Li, Jiatong and Lv, Zheqi and Yuan, Shaozu and Wang, Peiying and
             Zhang, An and Wang, Jiawei and Chua, Tat-Seng},
  journal = {arXiv preprint arXiv:XXXX.XXXXX},
  year    = {2026}
}

Copy BibTeX