73. Hermes Agent

The lesson

Hermes Agent is a self-improving, terminal-native autonomous AI agent built by Nous Research, the lab behind the Hermes, Nomos, and Psyche model families. Unlike a single-turn chatbot, it is designed to run multi-step tasks, use tools, remember context across sessions, and improve its own procedures over time.

The agent is built around a core orchestrator called AIAgent. Entry points such as the terminal, TUI, messaging gateways, and cron scheduler all feed into this orchestrator. Inside the loop, a prompt builder assembles identity, skills, memory, context files, and conversation history; a provider runtime handles model selection and API calls; a tool dispatch system executes tools; and session storage keeps everything searchable.

Hermes Agent ships with more than 70 built-in tools grouped into toolsets. Common toolsets include terminal, file, web, browser, memory, skills, delegation, cronjob, and vision. You enable only the toolsets you need, which reduces context use and limits what the agent can touch.

Persistent memory is central to Hermes Agent. MEMORY.md and USER.md are injected into every session, so durable facts and user preferences follow you across chats. Session history is stored in SQLite with FTS5 search, so the agent can look up past conversations. Project rules live in context files such as AGENTS.md, CLAUDE.md, .cursorrules, and .hermes.md.

Skills are reusable procedural knowledge documents. A skill is a markdown file with YAML frontmatter stored under ~/.hermes/skills/<category>/<skill>/SKILL.md. When installed, each skill becomes a slash command. If you find yourself explaining the same five-step workflow repeatedly, convert it into a skill.

Hermes Agent supports many providers: Nous Portal, OpenRouter, OpenAI, Anthropic, Google Gemini, z.ai/GLM, Kimi/Moonshot, DeepSeek, Ollama, LM Studio, vLLM, SGLang, and any custom OpenAI-compatible endpoint. The main model needs at least 64,000 tokens of context. Auxiliary models can handle vision, compression, title generation, and other side tasks.

Prompt engineering for an agent differs from single-turn chat. You must tell the agent what to do, when to use tools, and how to format results. Be specific, provide file paths and error messages, and let the agent iterate. For example, instead of 'fix the code,' say 'fix the TypeError in api/handlers.py line 47 where process_request() receives None from parse_body().'

Common workflows include daily briefing bots, GitHub PR review agents, parallel research with subagent delegation, code review and fix loops, and memory-driven project conventions. Cron jobs let you schedule recurring tasks and deliver results to Telegram, Discord, Slack, WhatsApp, email, and other messaging platforms.

Safety is your responsibility. Use Docker or Daytona for untrusted code, review dangerous commands before approving 'always,' and keep messaging gateway allowlists tight. Remember that the agent can take real actions: it can write files, run commands, and send messages, so keep a human in the loop for high-stakes decisions.

Troubleshooting: if hermes is not found, reload your shell PATH. If replies are empty, re-run hermes model or hermes doctor. If context length is exceeded, use /compress or start a new session. If subagents lose context, remember that delegate_task needs the full goal and context spelled out because child agents start with no memory.

Check yourself