Proposal: Integrate SIA (Self-Improving AI with Harness & Weight Updates) as an advanced harness evolution capability / skill

[jinsoo]

Hi Hermes team and community,

I’ve been following both the Hermes Agent project and the recently released SIA framework (https://github.com/hexo-ai/sia) with great interest.

SIA introduces a clean self-improving loop using three specialized agents (Meta-Agent → Target Agent → Feedback/Improvement Agent). The Feedback Agent analyzes execution logs and iteratively rewrites the target_agent.py (the harness) to better solve a given task. It has shown strong results on benchmarks like LawBench (+56.6% over prior SOTA in some configurations) by evolving the agent’s code structure rather than just prompts.

Why this might be relevant to Hermes

Hermes already has a powerful skill system and is actively exploring evolutionary self-improvement (via the hermes-agent-self-evolution repo, DSPy + GEPA, Darwinian Evolver ideas, etc.).

SIA’s approach feels conceptually aligned but complementary:

• It focuses on code-level harness evolution (tool dispatch logic, answer parsing, retry strategies, decision flow) through an explicit agentic loop.
• Hermes’ current evolution work is excellent at optimizing skills, prompts, and structured search. SIA adds a more agentic, log-driven code rewriting dimension.

For complex or long-horizon tasks, being able to automatically evolve a task-specific harness could be a very powerful addition.

Possible integration directions

I see two potential levels of integration:

1. Lightweight approach (easier to start)
   Create a Hermes Skill (e.g., SIAHarnessOptimizer) that can call SIA’s orchestrator for difficult tasks. Hermes would pass the task description, and SIA would return an improved harness/agent. The improved harness could then be converted into or saved as a reusable Hermes skill.

2. Deeper integration (more ambitious)
   Incorporate SIA-style meta-loop thinking into Hermes’ self-evolution pipeline. For example, when evolving skills for complex tasks, optionally run a short SIA-style improvement loop (log analysis → code/harness rewrite) alongside or instead of pure DSPy/GEPA search.

Benefits I see

• Stronger performance on tasks that require sophisticated tool use, parsing, or multi-step reasoning.
• A new “meta-skill” that can improve other skills/agents over generations.
• Synergy with Hermes’ existing persistent memory and multi-agent delegation features.

SIA is still very new (released May 28), and its current open-source version focuses primarily on harness evolution (weight updates are discussed in the paper but not yet in the repo). This might actually make it easier to integrate as a focused capability.

Would the team be interested in exploring this direction? I’d be happy to help draft a more detailed design doc, create a simple wrapper skill prototype, or discuss how SIA’s loop could best complement the existing DSPy + GEPA evolution work.

Thanks for all the great work on Hermes — really excited about where the self-improving agent space is heading.

Best regards,

---

References (you can add at the bottom or as links in the issue):

• SIA GitHub: https://github.com/hexo-ai/sia
• SIA Paper: https://arxiv.org/abs/2605.27276
• Hermes Self-Evolution repo: https://github.com/NousResearch/hermes-agent-self-evolution

---