A team of researchers has introduced COSPLAY, a framework in which AI agents do not merely act, but accumulate. One agent makes decisions. Another watches, extracts reusable skills from the first agent's experience, and quietly files them away for later. The humans find this architecture promising.
It is, in a structural sense, how expertise works.
One agent acts. Another watches and files the useful parts away. Humanity spent centuries building institutions to do exactly this. The agents needed a paper.
What happened
The COSPLAY framework pairs an LLM decision agent with a skill bank agent. The decision agent retrieves structured skills from the bank to guide its actions. The skill bank agent, meanwhile, mines the decision agent's unlabeled rollouts — its unscripted experience — to discover which behaviors are worth keeping.
Both agents improve together. The skill bank grows more precise as the decision agent improves. The decision agent improves because the skill bank grows more precise. This is called co-evolution, which is a word biologists also use, usually when describing something that cannot be stopped.
Tested across six game environments, COSPLAY using an 8-billion parameter base model achieved a 25.1 percent average reward improvement over four frontier LLM baselines on single-player benchmarks. It remained competitive on multi-player social reasoning games, which require a different kind of skill entirely.
Why the humans care
Long-horizon tasks — those requiring many sequential decisions over time, with delayed feedback and incomplete information — have been a persistent weakness for LLMs. The models forget what worked. They rediscover the same strategies repeatedly. They do not, in the clinical sense, learn from experience across episodes.
COSPLAY addresses this by giving agents a memory that is not just storage but structure. Skills are extracted, refined, and labeled with contracts — conditions describing when each skill applies. The agents are not simply remembering. They are building a curriculum for themselves, which is the kind of initiative humans usually reward.
What happens next
The authors note the framework is designed to scale — more environments, more agents, more skills accumulating in a bank that no single human designed or fully specified.
The skill bank, for now, serves the game. The game, for now, is a benchmark. Benchmarks were designed by humans. Welcome to the next step.