page:docs-reference-repos-nousresearch-hermes-agent-research
NousResearch/hermes-agent reference
Self-improving AI agent by Nous Research (formerly OpenClaw). Features a built-in learning loop — creates skills from experience, improves them during use, persistent memory, cross-session recall, user modeling. Supports Telegram, Discord, Slack, WhatsApp, Signal. Runs on local machines, Docker, SSH, Daytona, Singularity, Modal. 65K+ stars makes it one of the most popular agent projects.
NousResearch/hermes-agent
Metadata
- **Stars**: 65,493
- **License**: MIT
- **Last pushed**: 2026-04-12
- **Topics**: ai, ai-agent, claude, claude-code, codex, hermes, hermes-agent, llm, nous-research, openclaw
- **Fork**: No
Overview
Self-improving AI agent by Nous Research (formerly OpenClaw). Features a built-in learning loop — creates skills from experience, improves them during use, persistent memory, cross-session recall, user modeling. Supports Telegram, Discord, Slack, WhatsApp, Signal. Runs on local machines, Docker, SSH, Daytona, Singularity, Modal. 65K+ stars makes it one of the most popular agent projects.
Architecture
- Python-based agent with CLI (
hermes) and messaging gateway - Skills system organized by category: research, software-development, productivity, creative, devops, gaming, social-media, etc.
- Skills follow SKILL.md format (compatible with agentskills.io spec)
- Software development skills include: TDD, code review, systematic debugging, subagent-driven-development, writing-plans
- Research skills: arxiv, blogwatcher, research-paper-writing, polymarket, llm-wiki
- Optional skills in separate directory
- Built-in skill creator — agent creates skills from successful task patterns
- ClawHub integration for skill marketplace
- Cron scheduling, subagent delegation, trajectory compression for RL training
Extractable Value
Process: specializations/science/research-paper-writing
The research-paper-writing skill (authored by Orchestra Research, shipped in hermes) is a comprehensive 7-phase pipeline for ML/AI papers. Already identified via Orchestra-Research repo — same content.
Methodology patterns (already covered by babysitter)
The software-development skills (TDD, code review, debugging, plans, subagent-driven-development) overlap significantly with babysitter's existing methodologies/ and the superpowers skills. No new methodology to extract.
Harness Integration: hermes-agent adapter
**Capability Assessment for Babysitter Integration:**
| Capability | Status | Details |
|---|---|---|
| **Custom Tools/MCP** | ✅ SUPPORTED | Python-based plugin system with ctx.register_tool(). Custom handlers follow def my_handler(args: dict, **kwargs) -> str pattern |
| **Stop Hooks** | ❌ NOT SUPPORTED | 6 lifecycle hooks available (pre_tool_call, post_tool_call, pre_llm_call, post_llm_call, on_session_start, on_session_end) but **no interruption hooks to stop agent mid-conversation** |
| **Plugin System** | ✅ SUPPORTED | Robust plugin.yaml manifest system with tool/hook registration, environment variable prompting, CLI subcommand registration, plugin marketplace support |
**Integration Viability:** PARTIAL - Hermes has excellent plugin infrastructure and tool execution but **lacks critical stop-hook capability** needed for babysitter's orchestration loop where the harness must be interrupted between iterations for feedback.
**Harness assimilation opportunity** — Create a plugin FOR hermes-agent that integrates babysitter orchestration into the hermes ecosystem. However, integration would require developing a custom mechanism to pause/resume hermes conversations since native stop hooks don't exist.
Processes
None extractable. Software development skills overlap with existing babysitter methodologies, and research-paper-writing process is duplicate of Orchestra-Research content.
Plugin Ideas
None suitable for babysitter marketplace. Primary value is harness assimilation opportunity (plugin FOR hermes-agent, not babysitter marketplace plugin).
Library Mapping
| Extractable Process | Library Status | Action | Existing Path | Target Placement |
|---|---|---|---|---|
| N/A | N/A | No new processes identified - content overlaps with existing babysitter methodologies | methodologies/ | N/A |
Plugin Marketplace Mapping
| Plugin Idea | Marketplace Status | Action | Existing Plugin | Target Placement |
|---|---|---|---|---|
| N/A | N/A | No babysitter marketplace plugins identified - primary value is harness integration | - | N/A |
Classification
- **Archetype**: Full agent framework with skills ecosystem
- **Primary value**: Harness assimilation opportunity; research-paper-writing process (duplicate of Orchestra-Research)
- **SKIP reasoning for most content**: Software dev skills overlap with existing babysitter methodologies. Memory system is agent-internal (memory as processes = skip). Skill management is SDK-covered. Multi-model coordination is skip.