EvoGroup: A Self-Evolving Harness for Group Agents in Shared Contexts
EvoGroup treats an agent harness as the object of evolution rather than hand-designing group-memory behavior in advance. Starting from six basic tools, it records tasks over a shared context, prioritizes confident-correct, confident-wrong, and uncertain trajectories, then uses adaptive progressive disclosure and bucketed analysis to diagnose failures before revising memory, tools, middleware, prompts, and skills through representation, operation, intervention, and policy interfaces. On the EverMemBench held-out split, DeepSeek-V4-Flash improves from 71.07% to 89.58% pass@1, exceeding Codex at 82.62% and Claude Code at 85.83%. The frozen harness shows partial transfer on GroupMemBench, while model-dependent gains and revision regressions remain important limitations.