Hermes Agent is an open-source AI agent released by the New York–based lab Nous Research in February 2026. Since its launch it has briefly overtaken the popular OpenClaw on a daily-token leaderboard compiled by the AI-model platform OpenRouter. Some users have complained that Hermes Agent uses tokens less efficiently, but its capability to define and refine new skills — specialized instructions, workflows or domain knowledge — highlights self-improvement as a core agentic capability.
What it does and how it runs
Hermes Agent’s capabilities largely overlap with OpenClaw’s, but it differs mainly in its memory architecture and its automatic skill-building workflow. The agent can run locally or in the cloud, supports a wide variety of large language models, and integrates with roughly 20 messaging services. Using a model that runs locally (or one that generates new access tokens after a browser sign-in) allows operation without storing a persistent API key.
It also integrates with development environments via the Agent Communication Protocol.
The agentic loop
Hermes Agent follows an agentic loop similar to OpenClaw’s:
- The agent assembles a prompt from its defined personality, instructions, tools, skills, memory, knowledge about the user, and the conversation history including the most recent message.
- If the prompt exceeds the LLM’s input limit, the agent asks the LLM to summarize older messages to reduce size.
- It sends the assembled prompt to the LLM and either calls a tool, invokes a skill, or returns a response to the user.
- If it calls a tool or skill, that call executes and may produce another tool/skill call or a user-facing response. The cycle repeats until the model generates a final response for the user.
Skills and automatic skill creation
Hermes Agent uses the standard SKILL.md format for instruction files that tell the agent how to perform tasks — for example, running bash scripts, searching the web or files, querying databases, and so on. The agent ships with built-in skills, and additional skills are available from the Skills Hub (which is currently far smaller than OpenClaw’s large, crowd-sourced skill library).
In contrast to many agents, Hermes Agent can create new skills automatically. When the agent works on a problem for a long time or fixes an error and judges the task complete, it calls a tool to create a new skill. To prevent proliferation of agent-generated skills, a background system called Curator (i) archives any skill that has not been used in over 90 days by moving it to a separate folder, and (ii) uses an LLM to decide whether each skill should be kept as is, merged with other skills, or archived.
Memory
Hermes Agent maintains two general memory files that are added to prompts: one records user preferences, and the other captures workflows and lessons learned. A built-in memory tool handles additions. Before adding a memory, the agent checks whether it’s worth adding (for example, it won’t add a memory if a similar one already exists or the content is too vague) and which of the two files to use. If adding a memory would exceed a preset file length, the agent examines the relevant memory file and merges related entries.
The agent also keeps a searchable database of conversations accessible through a separate tool. Additionally, Hermes Agent can leverage external memory providers such as Honcho, which analyzes the user’s identity after every message to infer preferences, goals, and patterns.
Persistent goal tracking
Users can specify a goal in a message. After generating a response, Hermes Agent calls a judge model to evaluate whether the goal was achieved; if not, it continues working. The loop continues until the judge model determines the goal is completed or the agent reaches a maximum number of turns. Similar functionality is available from Anthropic Claude Code, OpenAI Codex, and OpenClaw (via a plugin).
Background and why this matters
Agentic capabilities emerged as large language models gained the ability to plan multi-step actions, reflect on earlier outputs, and use external tools to act online. Coding agents like Anthropic’s Claude Code and OpenAI’s Codex gained traction among developers in 2025 and helped build momentum for more-autonomous AI systems. In early 2026, OpenClaw became an open-source phenomenon with a personal agent that ran continuously, executed online tasks, and interacted through messaging platforms such as WhatsApp and Telegram; its inventor later joined OpenAI.
OpenClaw’s popularity and the security issues that surfaced at launch spurred a wave of “Claw”-like agents, including Hermes Agent in February 2026. Interest accelerated in late April and May as successive releases made the agent easier to use and its self-improving behaviors more robust.
Overall, general-purpose agents are rapidly expanding the scope of AI-driven capabilities. While a typical feature set is coalescing, new features continue to appear. Hermes Agent’s more sophisticated memory and its ability to turn successful behaviors into reusable skills point toward a transition from stateless AI assistants to agents that accumulate experience, adapt to users, and automate ongoing work beyond isolated tasks.
Final note
Open-source agents that are not tied to a single LLM, messaging platform, or skill format can be particularly valuable because they work within familiar communication channels and can leverage the best available models within their chosen harnesses.
