Signal. Nous Research shipped Hermes Agent v0.18.0 — “The Judgement Release” — and quietly moved the interesting frontier from which model to which topology. The pitch reads flat: one agent, one memory, six messaging surfaces plus the CLI (Telegram, Discord, Slack, WhatsApp, Signal, Email). But read Nous Research’s own feature list as a systems diagram, not a marketing sheet, and something sharper appears — a persistent assistant that connects across platforms, remembers your projects, auto-generates skills, schedules unattended jobs, delegates to isolated subagents, and runs across five backends: local, Docker, SSH, Singularity, and Modal. That last line is the whole story.

System. Most “AI agent” projects are a chat box bolted to an API key. Hermes is built the other way round — as an execution substrate with a gateway in front and namespace isolation underneath. The single-memory-many-surfaces model means your dependency graph collapses to one node you actually own, instead of six half-integrations you don’t. From the cartographer’s desk that is the correct move: reduce the number of things that can independently fail. The reason this is possible in 2026 and wasn’t in 2022 is the backend abstraction — Singularity and Modal weren’t casually cloneable runtimes for a hobby repo until containers and serverless GPU became boring. Boring is good. Boring is what you build on.

←TODAY: An MIT-licensed agent runtime clones in one afternoon and runs sandboxed in Docker on a studio laptop. →3012: The office that survives is the one whose automation ran on infrastructure it could redraw from memory. Fulcrum: Owning the topology — not the model weights — is what keeps the lights on when a vendor endpoint goes dark.

The Tool: The protagonist here is Hermes Agent by Nous Research, and its browser companion hermes-webui by nesquena — a Python-plus-vanilla-JavaScript front end that, per its Hermes Atlas card, carries an MIT license, sits at 15.6K GitHub stars, and was last updated 2026-07-08. The Atlas overview claims full functional parity with the Hermes CLI through a three-panel layout — chat sessions, workspace files, agent settings — with no build step and no framework tax. For a computational designer that combination (real terminal, real Python RPC, no npm swamp) is worth an afternoon.

Setup:

git clone https://github.com/nesquena/hermes-webui.git
cd hermes-webui
python -m venv .venv && source .venv/bin/activate
pip install -e .
hermes-webui --backend docker --port 8080
# open http://127.0.0.1:8080 — the three-panel UI proves the install

First steps:

1. In the settings panel, point the agent at your model of choice and pick the docker backend — never local for the first run. You want the agent’s terminal inside a container, not on your machine.
2. In the chat panel, ask it to read one folder of a real project (a Grasshopper definition export, a set of PDFs, a spec). Watch it auto-generate a “skill” — that persisted memory, feature #2 on the Nous Research list, is the point of the whole exercise.
3. Use the Schedule feature to set one natural-language recurring job: “every morning, summarise new files in this workspace.” It runs unattended through the gateway.

Street. On a working architect’s desk this replaces the swivel-chair problem: the tender lives in Email, the site photos in WhatsApp, the model coordination in Slack, the code in the CLI. Hermes’ one-memory design means the agent that read Monday’s Bauleitung note is the same agent that answers the Discord question about it — no re-briefing, no lost context. That is a real reduction in coordination overhead, and it is measurable the moment you stop copy-pasting between four apps.

Atelier: For a Swiss studio the interesting reach is isolation, not chat. The Delegate feature (#4 in the changelog) spins up subagents with their own terminals and Python RPC scripts at zero context cost — exactly the shape you want for a repeatable BEP-drafting or quantity-takeoff pipeline that must not see your whole workspace. Monday move: clone hermes-webui onto one machine, run it exclusively on the Docker (or Singularity) backend, and set a single written office policy — agents never run on the local backend. One line of governance now saves you the incident report later.

Hack: Before you trust any agent runtime with a live terminal, trace what it actually pulls in — the real dependency graph, not the README’s architecture picture. This is the cartographer’s reflex: find the third single-point-of-failure you didn’t know you had.

pip install pipdeptree
pipdeptree -p hermes-webui --graph-output svg > hermes_deps.svg
pipdeptree -p hermes-webui | grep -c "=="   # true count of transitive deps

The SVG is your map; the integer is your risk surface. If that number surprises you, you learned something before the agent ran a single command.

Move. We did not run out of compute in my time — we ran out of intact bandwidth, intact cooling, and people who remembered how the old pipeline worked. Single points of failure stay quiet until the morning they aren’t. Clone the repo, run it sandboxed on one of the five backends, and draw the dependency graph before you wire it to anything that matters.

Learn-it:

• Repo card: nesquena/hermes-webui on Hermes Atlas — 15.6K stars, MIT, Python source.
• Home / docs: hermes-agent.nousresearch.com — the six-feature overview.
• Changelog: Nous Research — v0.18.0 “The Judgement Release”.
• Cheat sheet: A community Hermes Agent slash-command reference.
• PAZ note: pair this with PAZ’s dependency-graph exercise — the topology you can redraw from memory is the one that survives.