Hermes Self-Evolving Collaboration System

Problem Background

Agents executing complex work can over-answer simple questions, blur multi-step projects into one long response, and replace missing evidence with hedges such as "maybe" or "probably". The issue described in the skill is not a lack of model capability, but a missing behavioral protocol: when to answer directly, when to enter project-management mode, when to delegate to subagents, when to persist memory, and when to compress context.

Core Mechanisms

Hermes Self-Evolving Collaboration System turns behavior into layered rules:

• Task routing: single-step or single-file work gets direct answers; multi-step, multi-file work, or 3+ tool calls triggers PM mode.
• Subagent delegation: S/A tasks can use parallel delegate_task; B tasks delegate only independent subtasks; C/D tasks avoid delegation to limit context-isolation cost.
• Learning loop: corrections, dissatisfaction, new rules, or reusable workflows trigger memory or skill_manage, turning experience into memories, skills, or patches.
• Context compression: light, normal, heavy, and emergency levels summarize or trim history while preserving decisions, rules, preferences, and task goals.
• Quantified self-review: metrics such as verification rate, post-task review rate, and repeated-error count constrain behavior rather than only producing conclusions.

Boundaries and Notes

It fits multi-turn agent workloads that need cross-session memory, skill reuse, and auditable retrospectives. For one-shot Q&A, latency-sensitive tasks, or workflows where automatic history compression is unsafe, disable PM mode and automatic compression, keeping only concise direct-answer rules.