Hermes Skill Forge 🔨

Self-evolving agent that turns repeated tasks into reusable Skills - autonomously.

Built for the NousResearch “Show us what Hermes Agent can do” hackathon.

What It Does

Hermes Skill Forge watches itself complete tasks, identifies reusable patterns, writes a Skill, tests it in a Python sandbox, refines it until quality >= 0.8, then publishes it to the agentskills.io marketplace.

The more it works, the smarter it gets.

Architecture

flowchart TD
    A([🔔 Repeated Task Detected]) --> B
    B[🔍 OBSERVESearch memory + marketplaceAvoid duplication] --> C
    C[💡 ABSTRACTExtract reusable patternDefine inputs & outputs] --> D
    D[📝 WRITECreate SKILL.mdWith YAML frontmatter] --> E
    E[🧪 TESTRun 3 test cases in sandboxHappy · Edge · Error] --> F
    F{quality_score >= 0.8?}
    F -- Yes --> G[🚀 PUBLISHSimulate PR toagentskills.io]
    F -- No  --> H[🔧 REFINEFix failuresRewrite skill]
    H --> E
    G --> I([🧠 UPDATE MEMORYSkill stored foreverNever duplicated])

    style A fill:#c0392b,color:#fff
    style G fill:#27ae60,color:#fff
    style F fill:#e67e22,color:#fff
    style I fill:#8e44ad,color:#fff

Hermes Features Used

Reward Function

pie title Skill Forge Reward Components
    "Skill Written - SKILL.md created?" : 30
    "Tests Executed - Real Python run?" : 25
    "Quality Score - Achieved >= 0.8?" : 20
    "Published - Passed quality gate?" : 10
    "Searched First - No duplication?" : 10
    "Documentation - Clear + examples?" : 5

Quick Start

pip install openai rich
set OPENROUTER_API_KEY=sk-or-...

python demo/demo_skill_forge.py --task web-summarizer
python demo/demo_skill_forge.py --task log-analyzer
python demo/demo_skill_forge.py --task code-reviewer

Demo Output (web-summarizer)

🧠 search_memory     → Memory is empty.
🔍 search_skills     → No matching skills found.
📝 write_file        → Written SKILL.md to ~/.hermes/skills/web-summarizer/
🧪 execute_code      → HAPPY ['sentence one', 'sentence two', 'sentence three']
🧪 execute_code      → EDGE  ['No content to summarize']
🧪 execute_code      → ERROR ValueError (handled gracefully)
[AUTO EVALUATOR]     → quality_score=1.00 ✓ - calling publish_skill
🚀 publish_skill     → Published! PR: github.com/NousResearch/agentskills/pull/1890

Quality score: 1.00 | Skills written: 1 | Published: 1

Demo Scenarios

Project Structure

graph LR
    A[hermes-skill-marketplace] --> B[skills/]
    A --> C[environments/]
    A --> D[demo/]
    A --> E[tests/]
    A --> F[docs/]

    B --> B1[skill-forge/SKILL.mdAgent playbook]
    C --> C1[skill_forge_env.pyAtropos RL environment]
    C --> C2[skill_forge_config.yamlTraining config]
    D --> D1[demo_skill_forge.pyStandalone demo]
    E --> E1[test_skill_forge_env.pyPytest suite]

    style B1 fill:#27ae60,color:#fff
    style C1 fill:#8e44ad,color:#fff
    style D1 fill:#e67e22,color:#fff

Running Tests

python -m pytest tests/ -v
# or without pytest:
python -c "from environments.skill_forge_env import smoke_test; smoke_test()"

Why This Wins

Most agents do a task and stop. Hermes Skill Forge does a task and makes itself better at that task forever.

Every skill it publishes becomes available to every other Hermes agent. That’s not automation - it’s self-directed skill acquisition at community scale.