self-improvement-operating-loop

name	self-improvement-operating-loop
description	Evolutionary self-improvement loop for Hermes Agent using DSPy + GEPA. Evolve skills, prompts, and tool descriptions through reflective genetic search.
version	1
author	NousResearch / Marco
metadata	{"hermes":{"tags":["self-improvement","optimization","dspy","skills","evolution"],"category":"automation","skill_dir":"/home/marco/hermes-agent-self-evolution"}}

Self-Improvement Operating Loop

Automated skill and prompt evolution for Hermes Agent using hermes-agent-self-evolution (DSPy + GEPA).

When to Use

• User says "improve this skill", "optimize my agent", or "self-evolve"
• A skill produces sub-optimal results and needs targeted improvement
• After discovering a recurring failure pattern in agent behavior
• Periodic maintenance: run on a schedule to keep skills sharp

How It Works

Read skill/prompt → Generate eval dataset → GEPA mutates
       ↑                                      ↓
Execution traces ←── Evaluate candidates ←── Constraint gates
       ↓
Best variant → PR against skill repo

No GPU needed. ~$2-10 per optimization run via API calls.

Setup (One-Time)

# Clone the evolution engine
git clone https://github.com/NousResearch/hermes-agent-self-evolution.git ~/hermes-agent-self-evolution
cd ~/hermes-agent-self-evolution
pip install -e ".[dev]"

# Point at your hermes-agent/skills repo
export HERMES_AGENT_REPO=~/.hermes/skills

Run Evolution

Evolve a specific skill

cd ~/hermes-agent-self-evolution

# With synthetic eval data (fast, no history needed)
python -m evolution.skills.evolve_skill \
    --skill github-code-review \
    --iterations 10 \
    --eval-source synthetic

# With real session history (more accurate, needs sessiondb)
python -m evolution.skills.evolve_skill \
    --skill github-code-review \
    --iterations 10 \
    --eval-source sessiondb

Evolve all skills at once

python -m evolution.skills.evolve_all \
    --iterations 5 \
    --eval-source synthetic

Check status

python generate_report.py --latest
ls ~/hermes-agent-self-evolution/reports/

Phase Roadmap

Phase	Target	Status
1	Skill files (SKILL.md)	✅ Done
2	Tool descriptions	🔲 Soon
3	System prompt sections	🔲 Soon
4	Tool implementation code	🔲 Soon
5	Fully automated loop	🔲 Soon

Constraint Gates

Every evolved variant must pass:

• pytest tests/ -q — 100% pass
• Skill size ≤ 15KB
• Tool description ≤ 500 chars
• Semantic preservation check
• Human PR review

Tips

• Run Phase 1 on skills that change frequently or underperform
• Use sessiondb source for skills that fail in specific contexts
• Keep HERMES_AGENT_REPO pointing to the skills directory, not the main agent repo
• Review generated PRs carefully — evolution can drift semantics
• Cost per run: ~$2-10 depending on iterations and eval source

Skills Added (2026-04-25)

• cron-conditional-execution — conditional cron predicates (price/state/time/shell)
• deep-research-iterative — iterative research loop with gap analysis
• multi-modal-output — video/animation/PDF generation pipeline
• decision-log — save reasoning before closing strategic threads
• context-cross-reference — multi-front project coordination pattern