Empty developer workstation at dawn with a whiteboard full of architecture sketches showing agent memory layers and skill nodes, symbolizing the self-improving Hermes Agent by Nous Research

Hermes Agent 2026: First Production-Ready Self-Improving Open-Source AI Agent

103,000 GitHub stars, MIT license, 118 skills

Nous Research released Hermes Agent v0.10 on 16 April 2026, a self-improving open-source agent. The project is growing faster than LangChain and AutoGen combined. For European decision-makers, the question is whether a credible alternative to proprietary agent platforms has now emerged.

Summary

Hermes Agent is an open-source AI agent framework from Nous Research, first released on 25 February 2026. Version 0.10.0 from 16 April 2026 bundles 118 skills, three-layer memory, and six messaging integrations. Seven weeks after release, the project passed 95,600 GitHub stars and has since exceeded 103,000. Its technical differentiator is GEPA, an ICLR 2026 Oral-accepted self-improvement mechanism that makes agents with 20 or more self-generated skills 40 percent faster on repeated tasks. Self-hosting on European infrastructure starts at 5 euros per month, and the MIT license prevents vendor lock-in. For European enterprises, the framework is a strategically interesting option; for enterprise-grade production use it still needs more maturity, audit logging, and clear governance.

What Hermes Agent is and why it matters now

Agentic AI describes AI systems that perform tasks autonomously and in multiple steps, make decisions, and interact with external tools without human intervention at every stage.

Nous Research released Hermes Agent v0.10.0 on 16 April 2026. Seven weeks after the first release on 25 February 2026, the project passed 95,600 GitHub stars and matched the historical growth curves of LangChain and AutoGen combined. For you as a decision-maker, this means Hermes Agent is no longer a research project. It is a framework you can run on your own server with production-ready self-improvement.

103k+
GitHub stars for Hermes Agent
118
skills in version 0.10.0
6
messaging integrations
MIT
license, no enterprise tier

The pace is striking. Release v0.8.0 on 8 April 2026 merged 209 pull requests and closed 82 issues. Only eight days later came v0.10.0 with the next layer of maturity: three memory tiers, six messaging gateways, and a closed learning loop. That is the release cadence of a well-funded research lab, not a hobby project.

25 February 2026

Initial release

Nous Research ships Hermes Agent on GitHub under MIT license, positioned as "the open-source agent that grows with you".

8 April 2026

Version 0.8.0 with GEPA self-evolution

209 merged PRs, 82 closed issues, native Google AI Studio integration, MCP OAuth 2.1 with PKCE, and the first production release with GEPA-based skill optimization.

16 April 2026

Version 0.10.0 with 118 skills

Current release with three-layer memory, six messaging integrations, and a closed learning loop that generates reusable skills from experience.

The three pillars: memory, skills, model agnosticism

Hermes Agent differentiates itself from established frameworks such as LangChain, CrewAI, or the Microsoft Agent Framework through three design decisions: persistent multi-layer memory, automatic skill generation from solved work, and strict model independence. The combination makes it the first open-source system with practical self-improvement.

Three-layer memory

Short-term context for the current session, persistent long-term conversations with FTS5 full-text search, and procedural skill memory with LLM summarization. Together the layers produce a model that can retrieve tasks from weeks ago.

Automatic skill creation

After complex tasks, the agent writes reusable skill documents on its own. These skills are refined during use and reinforced through periodic prompts in memory.

Model agnosticism

Supported providers include Nous Portal, OpenRouter with 200+ models, NVIDIA NIM, Xiaomi MiMo, z.ai GLM, Kimi Moonshot, MiniMax, Hugging Face, OpenAI, and custom endpoints. Live model switching works mid-session without a restart.

At the core is GEPA, a method introduced by Gupta and colleagues. GEPA stands for Genetic-Pareto and uses large language models to analyze complete execution traces. Instead of collapsing a reinforcement-learning reward into a single scalar, the system reads error messages, profiling data, and reasoning chains, and proposes targeted prompt improvements. The paper was accepted as an Oral at ICLR 2026.

GEPA lets open-source models such as gpt-oss-120b beat proprietary frontier models on enterprise tasks by about 3 percent, at 20 to 90 times lower cost.

Databricks Research ,
Speed gain on repeated tasks (20+ skills) 40%
Memory retrieval latency at 10,000 documents 10 ms
Scaling without dedicated vector database ~100k docs
Fixed overhead per API call (tool definitions etc.) 73%
Perspective

The 40 percent figure is a speed gain on domain-similar tasks, not a quality gain. GEPA makes the agent more efficient, not smarter. For routine pathways this matters; for new domains or principled decisions there is no automatic benefit.

Deployment, costs, and integrations

Hermes Agent runs on Linux, macOS, WSL2, and Android via Termux. Installation is a single shell command. For personal use, a VPS with 1 vCPU and 1 GB RAM at 5 euros per month (Hetzner, IONOS, or similar) is enough. For team deployments with always-on operation and multiple messaging gateways, independent reviews recommend 2 vCPU and 4 GB RAM.

Component Requirement Cost (estimate)
Framework license MIT, no usage caps 0 euros
VPS (personal use) 1 vCPU, 1 GB RAM from 5 euros per month
VPS (team, always-on) 2 vCPU, 4 GB RAM from 15 euros per month
LLM cost per complex task Budget models such as Claude Haiku 4.5, GPT-5.4 Mini, Hermes 4 70B about 0.30 US dollars
Messaging integrations Telegram, Discord, Slack, WhatsApp, Signal, Email, CLI included in the framework
Sandbox backends local, Docker, SSH, Singularity, Modal variable, depends on Modal usage

The sandbox choice is a meaningful differentiator. Hermes Agent separates tool-call execution into five backend options. For critical deployments, tools can run in isolated Docker containers or via SSH on secured remote hosts. Combined with the already integrated MCP OAuth 2.1 PKCE authentication standard and OSV malware scanning for MCP extensions (shipped in v0.8.0), the security architecture is unusually solid for a seven-week-old project.

209

PRs in v0.8

Merges in eight days demonstrate development velocity and community engagement.

15+

LLM providers

From Nous Portal via OpenRouter with 200 models down to local endpoints. No single-vendor dependency.

509

Contributors

Over 500 developers have already contributed. Strong participation from the open-source community.

Sovereignty

European perspective

Hermes Agent hits a nerve in the digital sovereignty debate. According to Bitkom, 99 percent of European enterprises want digital independence, but only 57 percent have an exit strategy for non-EU software. A self-hosted open-source agent on European infrastructure addresses exactly this gap. France's April 2026 move to order an exit from non-EU software makes the topic politically unavoidable.

Hermes Agent for EU sovereignty
Data stays on your European server, no telemetry
MIT license prevents vendor lock-in
Combines with local inference stacks (NVIDIA NIM, Hugging Face)
Optional memory layer Honcho is also self-hostable
MCP OAuth 2.1 with PKCE as the industry standard
Open compliance questions
EU AI Act transparency rules take effect 2 August 2026
Audit trails for agent decisions are not provided out of the box
Memory opacity makes evidence for regulators difficult
Skill vetting process must be built separately
GDPR documentation is not bundled with the project

Microsoft released the Agent Governance Toolkit on 2 April 2026 under MIT license. The toolkit covers the OWASP Agentic Top 10 and the EU AI Act, with 20 adapters for LangChain, CrewAI, Google ADK, and Microsoft Agent Framework. Even though Hermes Agent is not listed as an explicit adapter yet, the move signals that infrastructure for open-source governance is being built right now.

51%
European SMEs use or test AI (Bitkom 2026)
37%
plan to expand AI usage in 2026
99%
want digital independence (Bitkom)
57%
have an exit strategy in place

Hermes Agent vs OpenClaw and proprietary platforms

Hermes Agent competes directly with OpenClaw, the largest open-source agent by reach. OpenClaw has five and a half times more GitHub stars, but a very different security record. Choosing between the two comes down to ecosystem breadth versus learning depth.

Aspect Hermes Agent OpenClaw
GitHub stars 103,000+ 345,000+
Messaging integrations 6 (curated) 50+ (broad)
Skills 118 (curated) 2,857 ClawHub
Learning mechanism GEPA, autonomous skill creation Session-based, skill ecosystem
Memory architecture FTS5 plus LLM summarization, three layers Session-based with auto-notes
CVEs since January 2026 0 9 incl. CVE-2026-25253 (CVSS 8.8)
Supply chain incidents none known ClawHavoc with 341 malicious skills
Maturity young, fast iteration established, governance partners

OpenClaw's explosive growth exposed infrastructure readiness gaps. Microsoft and Cisco issued their own cautions about running it on standard infrastructure. Hermes Agent's smaller footprint and stricter vetting have avoided comparable public incidents so far.

The New Stack, April 2026

Compared to proprietary platforms such as SAP Joule , Microsoft Copilot Studio, or Salesforce Agentforce, Hermes Agent offers a very different value proposition. No license fees, no mandatory cloud migration, full data sovereignty, but also less deep integration into ERP or CRM systems. The choice depends on whether your use case lives in documented enterprise processes or in flexible, technically creative automation.

Challenges and risks

Hermes Agent is young and not ready for every enterprise scenario. Independent assessments from European mid-market reviewers in March and April 2026 are cautious. Too early for production use in regular operations, too little documentation for large rollouts, too little community support for critical paths.

API stability

Between v0.x releases, API stability is not guaranteed. Version pinning is explicitly recommended, and upgrade paths must be actively curated.

Memory opacity

What the agent has learned and stored is hard to audit. For GDPR and the EU AI Act, dedicated audit tooling is needed that Hermes does not ship with.

Attack surface

A permanently running systemd service with five messaging gateways is a sizable attack surface. The gateway itself must be locked down carefully.

Planning note: Self-learning is disabled by default and requires explicit configuration. That is a deliberate design choice, but it means the headline 40 percent speed gain only appears after active opt-in and a warm-up phase with your own domain data. Budget for this phase before presenting Hermes Agent to stakeholders.

One more point: Hermes Agent is not built as a code generation platform. For coding agent workflows , Claude Code, Aider, or Cursor are better fits. Hermes aims at persistent conversation, task automation, research, and routine orchestration. Skill quality varies on complex multi-phase tasks, especially when domain knowledge is thin.

Recommendation

What companies should do now

For most European enterprises, Hermes Agent is a strategic option in April 2026, not a production system. The right stance is a pilot under controlled conditions, enough to understand the self-improving approach and calibrate future decisions against proprietary platforms.

1

Run a scout

Assign a technology scout to track Hermes Agent as a reference for self-improving approaches. Read release notes, follow GitHub issues, assess community activity.

2

Launch an isolated pilot

Set up a pilot for developer productivity, research assistance, or routine automation. Do not integrate with production customer data processes.

3

Test EU self-hosting

Deploy Hermes Agent on a European VPS such as Hetzner or IONOS. Verify network isolation, firewall rules, and credential management.

4

Plan audit logging

Plan memory inspection and audit logging from the start. Do not retrofit. The EU AI Act requires documented decision logic.

5

Use model agnosticism

Use budget models for routine tasks, frontier models only for critical paths. Live switching allows fine-grained cost control without rewrites.

6

Run a security review

Conduct a security review of messaging gateway deployments before connecting real accounts. The systemd service is an extended attack surface.

Suggested sequence: set up the technology scout in May 2026, launch an isolated pilot in June, run the security review before connecting production accounts in July, and start the first productive routine automations from August 2026 alongside the start of the EU AI Act transparency rules.

Further reading

Frequently asked questions

What is Hermes Agent? +

Hermes Agent is an open-source AI agent framework from Nous Research, first released on 25 February 2026. The agent runs on your own server, stores conversations persistently, automatically creates new skills from solved tasks, and is freely usable under the MIT license. Version 0.10.0 from 16 April 2026 bundles 118 skills, three-layer memory, and six messaging integrations.

What makes Hermes Agent special? +

Hermes Agent is the first open-source agent with production-ready self-improvement based on GEPA. Agents with 20 or more self-generated skills are 40 percent faster on repeated tasks in the same domain, according to Nous Research. The project reached 95,600 GitHub stars in seven weeks, faster than LangChain and AutoGen combined. Add strict model agnosticism with 15+ LLM providers and a consistent MIT license without an enterprise tier.

What does Hermes Agent cost? +

The framework itself is free under MIT license. Self-hosting on a European VPS starts at 5 euros per month for personal use with 1 vCPU and 1 GB RAM. Team deployments with always-on operation need 2 vCPU and 4 GB RAM, available from about 15 euros per month. LLM costs are about 0.30 US dollars per complex task using budget models such as Claude Haiku 4.5, GPT-5.4 Mini, or Hermes 4 70B.

Is Hermes Agent ready for production use in the mid-market? +

Only to a limited extent. Documentation is incomplete, the community is young, and API stability between v0.x releases is not guaranteed. For pilot projects covering developer productivity, research assistance, or routine automation, Hermes Agent is a reasonable option. For critical ERP or customer-data processes, enterprises should currently stick with more established platforms such as SAP Joule or Microsoft Copilot Studio, or wait for a more mature release.

How does Hermes Agent fit the EU AI Act? +

Hermes Agent maps well to GDPR requirements through self-hosting on European infrastructure. For the EU AI Act starting 2 August 2026, additional audit logging, transparency documentation, and a skill vetting process are required. Microsoft released its Agent Governance Toolkit on 2 April 2026 under MIT license, which provides compliance patterns for open-source frameworks and covers the OWASP Agentic Top 10 as well as EU AI Act requirements.

How does Hermes Agent compare to OpenClaw? +

OpenClaw has more reach with 345,000 GitHub stars and over 50 messaging integrations, but faced nine CVEs in March 2026 including CVE-2026-25253 with CVSS 8.8, plus the ClawHavoc supply chain attack with 341 malicious skills. Hermes Agent is smaller with 6 messaging integrations and 118 curated skills, free of known CVEs so far, and more focused on learning architecture and security. OpenClaw wins on ecosystem breadth, Hermes on learning depth and security posture.

What is GEPA? +

GEPA stands for Genetic-Pareto and is a method for automatic prompt and skill optimization. Unlike reinforcement-learning approaches that condense execution traces into a single reward scalar, GEPA uses an LLM to read full traces with error messages, profiling data, and reasoning chains, then proposes targeted fixes. The paper was accepted as an Oral at ICLR 2026 and, according to Databricks Research, enables open-source models to beat proprietary frontier models on enterprise tasks at 20 to 90 times lower cost.