SAGE Shows AI Agents Learn Better by Watching Peers Than Solo Practice

The dominant paradigm for self-improving AI agents is essentially solitary: an agent attempts a task, gets feedback, reflects on what went wrong, and tries again. It is the machine equivalent of studying alone in a library. But in the real world, agents increasingly operate in ecosystems where other agents' strategies and outcomes are publicly visible. A new paper asks a deceptively simple question: does watching your peers actually make you better?

The answer, it turns out, is nuanced in ways that matter a great deal for how we build and deploy agentic systems. Researchers introduce SAGE (Social Agent Group Evolution), an evaluation framework designed specifically to measure whether shared peer experience produces improvements that solo self-improvement cannot.

The Gap Nobody Was Measuring

Self-improving agents have become a hot area of research. Frameworks like Reflexion store textual reflections in memory to improve future attempts, while more recent systems use reinforcement learning to generate their own training data through self-play. The unsolved problem -- the one that separates research curiosity from real economic value -- is whether an agent can get better at its job without a human redesigning it. Most of this work, however, evaluates agents in isolation.

Recent research focuses on enabling agents to improve autonomously without human supervision through self-evolution, with self-improvement training encompassing self-rewarding, self-play, and self-refinement mechanisms. But none of these frameworks ask: what happens when an agent can also watch what its peers are doing? SAGE is the first framework built to answer exactly that question, with a rigorous compute-matched experimental design.