Hermes Agent

Hermes Agent is for the person who wants to point an agent at a codebase and leave. The trait that decides it is skill creation: after a hard task it writes a skill for next time, so the second week is better than the first without you editing prompts. It drives models, not other coding agents, and reaches you on Telegram, Discord or Slack when it needs an answer.

Pick it if you like an agent that accumulates. Pick OpenClaw if you want a general assistant with a wider plugin surface, and Claude Code if you want a coding agent that does exactly what it did yesterday.

The risk is drift. Autonomous skill creation means the toolset in month three is not the one you installed in month one, and no review step is documented between a skill being written and being used. Run locally, the first execution option listed, those skills execute with your permissions. The documented Windows caveat that Defender flags the bundled uv.exe is a small thing, but it tells you how young the packaging is.

Pin the skills directory in git and read the diff weekly. What it does right: subagents are spawned isolated, so a parallel workstream cannot pollute the parent's context.

Two documented mechanisms carry the design. 1. Memory: session transcripts are indexed with FTS5 and summarised by the model for cross-session recall, with a separate user-modelling layer, so context is retrieved rather than stuffed. 2. Execution: the same tool calls dispatch to one of seven terminal backends, local, Docker, SSH, Singularity, Modal, Daytona or Vercel Sandbox, so isolation is a configuration choice rather than a rewrite.

No benchmark is published and the self-improvement claim is asserted, not measured; a before-and-after on a fixed task set would settle it. The observation: a memory built on full-text search is the one part that will not change when the model does.

The structure is clean: MIT agent as distribution, Nous Portal as revenue. Portal bundles 300-plus models with web search, image generation, text-to-speech and a cloud browser from $20 a month, and the free tier is the classic top of funnel. Nous also trains its own Hermes models, so the agent is a showroom for the lab, which is pricing power most wrappers lack.

Likely acquirer: an inference provider or a larger lab that wants both the models and the crowd. Likely pivot: the Portal becomes the product and the agent becomes a client. Position: long, with the caveat that lab economics decide the runway, not agent adoption.

The demo is an agent answering in Slack. Procurement: the gateway puts an agent inside the company Slack and email, and the built-in cron runs jobs overnight with delivery to any platform, so sixty engineers means sixty unattended schedulers touching sixty repositories. There is no SSO, no central log of what ran, and the vendor is a research lab, not a support organisation.

There is no headless CI mode. Onboarding is a curl script and a key, which is cheap. Not yet. Revisit when a team tier with central key issuance and a retention policy exists.

MIT and Python, so I can read the loop and patch it. Configuration lives under ~/.hermes and is set with hermes config set, model access takes any OpenAI-compatible endpoint, which means Ollama, vLLM or llama.cpp on my own box, and MCP servers attach through the documented feature page. Nothing here needs the Portal.

The --portal flag in hermes setup --portal is the only vendor-shaped path, and it is optional. A fork would keep the gateway, the memory and the backends and lose nothing that matters. Grudging respect for a lab that shipped the open half first.