yuhuajiang2002/harnessmm

HarnessMM is an agent-harness efficiency system: an actor authors reusable actions, receipt checks gate their reuse, and eligible actions execute through native tools while recording substituted callbacks. Its companion recursive self-improvement work uses task feedback to refine prompts, context, and adapter rules, with validation selecting frozen candidates before testing. The repository also contains host integrations, task-specific harnesses, learning methods, and evaluation/accounting modules. Edges follow README-described behavior or explicit source-index references; many implementation details are unsampled, so uncertain links are omitted.