Feature: Self-Evolving Agent Roles via Skill Clustering

Parent: #344 (Multi-Agent Architecture) Related: #376 (Debate Mode), #377 (Shared Memory)

Motivation

#344 defines a vision for multi-agent with 6 pre-defined archetypes (Coordinator, Researcher, Developer, Browser Agent, Reviewer, Synthesizer). This is powerful but static — roles are hand-crafted by developers.

What if roles emerged organically from actual usage patterns?

Hermes already has a skill system where agents create and refine skills from experience. As the skill library grows (40+ skills in a typical active installation, spanning devops, creative, finance, research, data-science, etc.), natural clusters form:

devops cluster:   kanban, cronjob, git, webhook, backup-restore
coding cluster:   github-pr, code-review, tdd, debugging, python-debugpy
research cluster: arxiv, web_search, web_extract, technical-idea-research
creative cluster: ascii-art, pixel-art, excalidraw, manim-video, comfyui
finance cluster:  beancount-bookkeeping, receipt-expense, repayment-reminder

Proposal

A periodic agent evolution loop that:

1. Scan — Periodically (daily/weekly) analyze the skill library: tags, descriptions, usage frequency, co-occurrence patterns
2. Cluster — Group skills into functional domains using tag overlap + embedding similarity
3. Evolve — For each cluster, auto-generate an agent role definition with:
   • Bound skill subset (role's toolset)
   • Auto-generated system prompt from skill descriptions
   • Role name derived from the cluster's functional domain
4. Adapt — New skills automatically re-assigned to the best-fit cluster on next scan. Stale clusters (no skill additions/usage in N days) are pruned.
5. Collaborate — Clusters that frequently appear together in tasks can form debate pairs (e.g., "coder + reviewer" → feeds into #376)

Example Output

After scanning a 40+ skill library, the system produces:

roles:
  - name: devops-agent
    skills: [kanban-orchestrator, cronjob, webhook-subscriptions, hermes-backup-restore, github-repo-management]
    system_prompt: "You are a DevOps specialist. Manage cron jobs, backups, webhooks, and repository operations."
    
  - name: coding-agent  
    skills: [github-pr-workflow, github-code-review, test-driven-development, systematic-debugging, python-debugpy]
    system_prompt: "You are a software engineer. Write, review, debug, and test code."
    
  - name: research-agent
    skills: [arxiv, web_search, web_extract, technical-idea-research, spike]
    system_prompt: "You are a research analyst. Search papers, extract information, explore ideas."

Integration with #344

This proposal complements #344 by providing a bottom-up, data-driven source of agent roles, while #344 provides the top-down orchestration infrastructure:

• Clustered roles → #344's Agent Role definitions
• Clustered tool subsets → #344's role-specific tool sets
• Co-occurring clusters → #376's debate pair candidates
• The workflow DAG in #344 can be populated with auto-discovered rather than hand-crafted roles

Implementation Notes

• Leverages existing skill YAML frontmatter (tags, description, category)
• skill_view() already provides structured metadata
• Could be a cron-triggered meta-skill or a new hermes roles evolve CLI command
• Role definitions could be stored alongside skills in ~/.hermes/skills/ with a .role.yaml extension
• No new infrastructure needed — reuses skill system + cron + existing tools

Prior Art

• CAMEL-AI's RolePlaying assigns personas to agents programmatically
• The idea of emergent specialization from task clustering is well-studied in multi-agent RL
• Hermes' own skill creation loop (agents create skills from experience) is the natural foundation

Open Questions

• Minimum skill count for meaningful clustering? (suggest ≥10)
• Re-cluster frequency? (suggest weekly, or on-demand via command)
• Should user have veto over auto-generated roles?
• How to handle skills that genuinely span multiple domains?