hermes - 💡(How to fix) Fix Feature Request: Integrate Microsoft SkillOpt for Self-Evolving Agent Skills

Feature Request: Integrate Microsoft SkillOpt for Self-Evolving Agent Skills

Summary

Microsoft recently released SkillOpt — a text-space optimizer that trains reusable natural-language skills for frozen LLM agents through trajectory-driven edits, validation-gated updates, and deployable best_skill.md artifacts.

This issue proposes exploring SkillOpt integration with Hermes to enable automated skill optimization without touching model weights.

What is SkillOpt?

• Paper: arXiv:2605.23904
• Project Page: https://microsoft.github.io/SkillOpt/
• License: MIT

Core concept: Train agent skills like you train neural networks — with epochs, (mini-)batchsize, learning rates, and validation gates — but without touching model weights.

Pipeline (6 stages):

1. Rollout — execute episodes with current skill
2. Reflect — analyze trajectories, generate patches
3. Aggregate — hierarchical merge of patches
4. Select — rank and select top edits
5. Update — apply edits to skill document
6. Evaluate — validate candidate skill, accept/reject

Output: best_skill.md — a deployable skill document

Why Integrate with Hermes?

Proposed Integration Paths

Option A: CLI Wrapper Skill (Low Effort)

Create a Hermes skill that wraps SkillOpt CLI:

hermes skillopt train --config configs/hermes/default.yaml --split_dir ./my_tasks
hermes skillopt import ./outputs/best_skill.md --name my-optimized-skill

Option B: Native Integration (Medium Effort)

• Add SkillOpt as optional dependency
• Provide hermes skill optimize <skill-name> command
• Auto-generate training data from Hermes session history
• Integrate with existing skill registry

Option C: Deep Integration (High Effort)

• Replace/adapt Hermes skill system to use SkillOpt's document-based skill representation
• Real-time skill optimization during agent execution
• Continuous learning from user feedback

Technical Compatibility

• ✅ Python 3.10+ (matches Hermes)
• ✅ Supports Azure OpenAI / OpenAI / Anthropic / local vLLM
• ✅ MIT license
• ⚠️ Requires JSON training data in specific format
• ⚠️ Currently benchmark-oriented (SearchQA, ALFWorld, etc.), needs adaptation for general tasks

Next Steps

1. Evaluate SkillOpt on a Hermes skill (e.g., github-pr-workflow or requesting-code-review)
2. Prototype Option A as a proof-of-concept
3. Gather feedback from maintainers and community

References

• SkillOpt GitHub
• SkillOpt Paper
• SkillOpt Demo Video

This is a research/experimental feature proposal. Happy to contribute a POC if there's interest!