01 // What Is Hermes Agent?
NousResearch is a well-regarded open-source AI lab that has spent the last two years building one of the most capable fine-tuned model families outside of the frontier labs. Their Hermes models — trained specifically for agentic behavior, structured tool-calling, and instruction adherence — have accumulated a serious developer following on Hugging Face. In February 2026, NousResearch formalized something that had been gestating in that model work: a full autonomous agent framework bearing the same name.
Hermes Agent launched publicly on February 26, 2026. It is worth being precise about the distinction: the Hermes models (Hermes 2 Pro, Hermes 3, Hermes 4) are the underlying LLMs — open-weight, downloadable, and independently useful. The Hermes Agent framework is the autonomous runtime built on top of them. You can run Hermes Agent against any backend — Claude, GPT-4o, or the Hermes models themselves — just as you can run OpenClaw against a range of providers.
The framework is MIT-licensed on GitHub and reached 20,400 stars and 2,500 forks within weeks of launch — making it one of the fastest-growing open-source agent projects of early 2026.
02 // Why It's Blowing Up Right Now
Three converging factors explain the momentum:
The 405B Hermes 4 model posted benchmark scores competitive with GPT-4o on reasoning tasks and beat closed models on RefusalBench — 57.1% vs GPT-4o's 17.67%. That drew significant developer attention to the entire Hermes ecosystem months before the framework launched.
The February release moved Hermes Agent from an experimental repo to a documented, versioned release with a dedicated docs site, Skills Hub, and messaging gateway — making it enterprise-usable for the first time.
The field is visibly moving from single-session agents toward agents that accumulate knowledge. Hermes is the most prominent open-source entry in this category — and notably, it ships a built-in OpenClaw migration tool, signalling a direct challenge to the incumbent.
230 contributors, 20k+ GitHub stars in under five weeks. The OpenClaw community forums and Reddit have seen a marked uptick in "should I switch?" threads since the February launch.
03 // The Hermes Model Lineage
Understanding the framework requires understanding the models that preceded it. NousResearch has been releasing Hermes fine-tunes since 2023, each iteration pushing harder on agentic capabilities. You do not need to use these models to run Hermes Agent, but they inform the framework's design choices — particularly around structured tool-calling and the function-calling XML format.
| Version | Base Model | Key Agentic Capability | Technical Report |
|---|---|---|---|
| Hermes 2 Pro | Mistral / Llama 2 | ChatML format; first serious function-calling fine-tune on open weights | — |
| Hermes 3 | Llama 3.1 · 8B / 70B / 405B | Standardized <tool_call> XML format; JSON schema adherence; structured outputs |
arXiv:2408.11857 |
| Hermes 4 | Llama 3.1 · 14B / 70B / 405B | Hybrid <think> reasoning toggle; Atropos RL training; 19B tokens, 5M samples |
arXiv:2508.18255 |
Hermes 4's hybrid reasoning deserves particular attention: the model can decide
on a per-request basis whether to emit a <think> block before
answering. Unlike pure chain-of-thought models, this makes inference faster on
simple tasks while preserving deep reasoning capacity for complex ones. The training
was done using NousResearch's
Atropos RL environment,
which is open-source — meaning developers can replicate the training pipeline.
04 // Core Architecture
Hermes Agent is built around four architectural pillars that distinguish it from most agent frameworks, including OpenClaw. Each one compounds the others — the memory system feeds the skill system, the skill system reduces reliance on raw context, and both are available across all execution backends.
Multi-Level Memory System
Per the
official memory documentation,
Hermes operates three layers in parallel. A MEMORY.md file (~800 tokens)
holds environmental facts and project context, injected frozen into every system prompt
for KV-cache efficiency. A USER.md profile (~500 tokens) accumulates
user preferences and working style over time. And a SQLite database at
~/.hermes/state.db stores all past sessions with FTS5 full-text search
and Honcho user
modeling — an AI-generated user understanding layer that runs cross-platform. All
memory writes are security-scanned for prompt injection and data exfiltration patterns
before being committed.
Skills System — Procedural Memory
After any task requiring five or more tool calls, Hermes Agent auto-generates a
Skill Document: a structured SKILL.md file with frontmatter
fields covering name, version, when to use, step-by-step procedure, pitfalls, and
verification steps. Skills are stored in ~/.hermes/skills/ and follow
the agentskills.io
open standard — meaning they are portable to any compatible agent, not locked to Hermes.
The agent can create, patch, and delete its own skills via a built-in
skill_manage tool.
Execution Backends
Six backends are supported: Local, Docker, SSH, Daytona (cloud dev environments), Singularity (HPC), and Modal (serverless, near-zero idle cost). This range — from embedded to fully serverless — makes Hermes viable across a wider deployment surface than most open-source agents.
Multi-Platform Gateway
Like OpenClaw, Hermes Agent supports CLI, Telegram, Discord, Slack, WhatsApp, and Signal. The key difference is cross-platform conversation continuity: a task started in the CLI can be resumed in Telegram. OpenClaw treats each channel as an independent integration.
05 // OpenClaw, Briefly
OpenClaw is the open-source personal AI assistant that ClawBeat was built to cover. Written in TypeScript, it runs an always-on agent that connects to WhatsApp, Telegram, Slack, Discord, Signal, iMessage, and others. Its architecture is infrastructure-first: a hub-spoke model where one agent instance manages multiple messaging contexts, with a configurable heartbeat scheduler, 52+ built-in skills, and application-level security via allowlists.
06 // Head-to-Head Comparison
The table below compares Hermes Agent and OpenClaw across the dimensions developers care most about. Sources are linked in the Sources section below.
| Dimension | Hermes Agent | OpenClaw |
|---|---|---|
| Philosophy | Agent-first — grows through experience | Infrastructure-first — hub-spoke multi-channel |
| Memory | Multi-level: session + persistent + SQLite FTS5 + Honcho user modeling | Per-assistant isolated storage; shared team context |
| Skill learning | Auto-generates SKILL.md after 5+ tool calls; agentskills.io portable | 52+ built-in skills; file-based precedence; no auto-generation |
| Channels | CLI, Telegram, Discord, Slack, WhatsApp, Signal + cross-platform continuity | WhatsApp, Telegram, Slack, Discord, Signal, iMessage, browser chat |
| Scheduling | Natural-language cron + parallel subagent workstreams | Heartbeat scheduler (configurable intervals) |
| Execution | 6 backends: Local → Docker → SSH → Daytona → Singularity → Modal | Managed (under 5 min) or self-hosted |
| Model breadth | 200+ models via OpenRouter | Core 6 + OpenRouter |
| Migration tooling | Built-in: auto-imports configs, memories, skills from OpenClaw | — |
| Telemetry | Zero — no usage data collected | Standard telemetry (opt-out available) |
| License | MIT | MIT |
07 // What Developers Should Know
Six technical details that are easy to miss but matter in practice:
If using Hermes models directly — not via the agent framework — tool calls must use the
hermes-function-calling XML/JSON schema, not OpenAI's
tool_choice format. NousResearch maintains adapters and complete examples
in the
Hermes-Function-Calling repo.
The agent framework handles this translation automatically.
The SKILL.md format is a structured markdown document with YAML frontmatter.
Developers can write skills for their specific use cases and drop them into
~/.hermes/skills/ before ever running the agent. This means a team
can pre-seed domain knowledge — deployment procedures, debugging runbooks,
API workflows — before onboarding.
Hermes Agent connects directly to any MCP server with tool filtering — meaning it plugs into the same tool ecosystem as Claude Code, Cursor, and other MCP-aware agents. If your team already has MCP servers deployed, Hermes can use them on day one.
Skills authored in Hermes' SKILL.md format are compatible with the agentskills.io open standard — designed to make skills portable across any compliant agent. This is early-stage but strategically important: if the standard gains traction, a skills ecosystem could emerge that benefits any agent that adopts it.
The reinforcement learning environment used to train Hermes 4 — Atropos — is publicly available with ~1,000 task-specific verifiers. Developers working on custom fine-tunes or research applications have access to the same pipeline NousResearch used. This is unusually transparent for a model lab.
Hermes Agent injects memory files as frozen snapshots at session start. This is a deliberate performance optimization: by keeping the system prompt prefix identical across sessions, the LLM's KV cache can be reused. The tradeoff is that mid-session memory updates do not affect the current context — they apply on next session load.
08 // Which One Is Right For You?
Both frameworks are MIT-licensed, actively maintained, and production-capable as of Q1 2026. The right choice depends primarily on what you need the agent to do — and whether its operating model should be breadth-first or depth-first.
- You want a personal research or development agent that improves on repeated task types
- Privacy is a hard requirement — zero telemetry, fully local by default
- You need flexible deployment (serverless, HPC, cloud dev environments)
- You want to experiment with or extend the underlying model via Atropos RL
- You are already on OpenClaw and want to migrate with context preserved
- Your workload involves complex, multi-step tasks that benefit from a skills library
- You are deploying for a team across multiple messaging channels
- You need the fastest path to a working assistant — under 5-minute managed setup
- iMessage integration is a requirement (Hermes Agent does not support it)
- You want a mature ecosystem with 52+ documented skills and active community plugins
- Your use case is reactive and multi-channel rather than a single persistent agent
09 // Sources
- 01Hermes Agent — Official site, NousResearchOfficial
- 02Persistent Memory — Hermes Agent DocumentationOfficial
- 03Skills System — Hermes Agent DocumentationOfficial
- 04NousResearch/hermes-agent — GitHubGitHub
- 05NousResearch/Hermes-Function-Calling — GitHubGitHub
- 06Hermes 4 Technical Report — arXiv:2508.18255Paper
- 07Hermes 3 Technical Report — arXiv:2408.11857Paper
- 08Hermes 4 405B — Hugging Face model cardModel
- 09Hermes 3 — NousResearch announcementOfficial
- 10Atropos RL training environment — NousResearch GitHubGitHub
- 11agentskills.io — Open skills portability standardStandard
- 12NousResearch Releases Hermes Agent — MarkTechPost, Feb 26 2026News
- 13Hermes 4: Hybrid Reasoning Models — MarkTechPost, Aug 2025News
- 14Hermes Agent: A Self-Improving AI Agent That Runs Anywhere — DEV CommunityAnalysis
- 15openclaw/openclaw — GitHubGitHub