Hermes Agent · Nous Research · 2026

How the Hermes Agent Works — and Its Components

A mental model for Nous Research's open-source (MIT) Hermes agent: one agent, one memory, every surface. We'll walk its six capabilities — Connect, Remember, Schedule, Delegate, Search, Experiment — then trace a single request end to end, through persistent skill-learning memory, isolated subagents, the scheduling gateway, and five hardened execution backends.

0 / 6 sections
1
One Agent, One Memory, Every Surface
The core idea: Hermes lives everywhere — and it's always the same agent behind every channel

Hermes is "the agent that grows with you." Its founding idea is that you shouldn't have a different assistant in every app. Instead it lives everywhere — Telegram, Discord, Slack, WhatsApp, Signal, Email, the CLI, and a growing list of platforms — but every one of those surfaces is just a doorway into a single agent with one shared memory.

So a project you set up from the CLI is known when you message it on Slack; a fact it learned over email is available on WhatsApp. The surface changes; the agent and its memory don't.

Demo — message Hermes from any surface; notice it's the same agent + memory
🧠
One Hermes agent
one shared memory
Click any surface to send a message from it.
The demos on this page are illustrative simulations of the behaviour described on the Hermes site, built to make each concept tangible — not live calls to the real agent.
What does "one agent, one memory, every surface" actually mean for Hermes?
2
The Six Capabilities — The Component Map
Connect · Remember · Schedule · Delegate · Search · Experiment

Hermes is organized around six capabilities. Together they're the building blocks of everything it does — think of them as the agent's components. The rest of this page zooms into the four that explain how it works: Remember, Delegate, Schedule, and the Search/Experiment pairing.

Demo — click each capability to see what it does
Click a capability card.
Which Hermes capability covers isolated subagents with their own terminals and Python RPC scripts?
3
Remember: Persistent Memory That Auto-Generates Skills
It learns your projects, saves how it solved a problem, and never re-derives it

Memory is what makes Hermes "grow with you." It does three things: learns your projects, auto-generates skills, and never forgets how it solved a problem. The key move is that auto-generated skill: the first time it works out how to do something, it saves that recipe — so the next time, it reuses the skill instead of reasoning from scratch.

Demo — run a task twice. First run learns a skill; second run reuses it
Auto-generated skill library
No skills yet — run a task to teach Hermes one.
Pick a task. The first time it's slow (figuring it out); the second time it's instant (reused skill).
This is the difference between a stateless chatbot and an agent that compounds: every problem solved becomes reusable capability, so it gets faster and more capable the longer you use it.
Hermes is asked to do a task it already worked out last week. What does its persistent memory let it do?
4
Delegate: Isolated Subagents, Zero-Context-Cost
Each subagent gets its own conversation, terminal, and Python RPC — and returns only the result

Big jobs blow up an agent's context window. Hermes' answer is to delegate: it hands a chunk of work to an isolated subagent that has its own conversation, its own terminal, and its own Python RPC scripts. The subagent does the messy, token-heavy work in isolation and hands back only the result.

That's what "zero-context-cost pipelines" means — the main agent's context stays lean because all the intermediate thinking happened somewhere else, out of view.

Demo — delegate a job to three isolated subagents and watch the main context stay lean
0
Tokens if run inline in the main context
0
Tokens added to the main context when delegated
The main agent only needs the answer — not every step the subagents took to get there.
Why are Hermes subagents described as "zero-context-cost"?
5
Schedule: Unattended Work Through the Gateway
Describe a recurring task in plain English; the gateway runs it on its own, focused every time

You don't write cron syntax for Hermes — you describe what you want in natural language: "every morning, email me a sales briefing." Hermes turns that into a schedule and runs it unattended through the gateway — a process that keeps firing your reports, backups, and briefings even when you're offline, "focused every time."

Demo — pick a natural-language schedule, then fast-forward and watch the gateway fire it
Pick a schedule above to arm the gateway.
Why a gateway? Scheduled runs can't depend on your laptop being open or a chat window being focused. The gateway is the always-on process that owns the clock and launches the agent on time, on its own.
How does a scheduled Hermes job run while you're offline?
6
The Full Flow: Tools, Backends, and One Request End to End
Search (the toolset) + Experiment (the five backends) — assembled into a single run

The last two capabilities are where work actually gets done. Search is the toolset the agent reaches for; Experiment is the set of places that work can run. Switch between them, then play the full flow to see how every component from this page snaps together.

Search gives the agent hands and senses. Backed by access to 300+ models via the Nous Portal, it can pick the right model for the job — that's the "multi-model reasoning."

🌐 Web search 🖱️ Browser automation 👁️ Vision 🎨 Image generation 🔊 Text-to-speech 🧠 Multi-model reasoning

Experiment is where code and tools execute — five interchangeable backends, all with container hardening and namespace isolation so a task can't reach beyond its sandbox.

💻 Local 📦 Docker 🔗 SSH 🧊 Singularity ☁️ Modal
Local for quick work, Docker/Singularity for reproducible isolated runs, SSH to use a remote box, Modal for serverless scale — same agent, swappable execution.

One request, all components

Hover any component for its job, then hit Play full flow to trace a real request through the whole system.

Demo — "research competitor pricing and post a summary to Slack"
📨
Surface
Slack message
The request enters from any connected surface — Slack, Telegram, CLI, email. Hermes treats them all as one inbox into a single agent (Section 1).
🧠
Agent Core
reasons & plans
The core reasons over the request with access to 300+ models via the Nous Portal (multi-model reasoning), then decides what to recall, which tools to call, and what to delegate.
💾
Memory
skills & projects
Persistent memory + auto-generated skills. The core checks whether it has solved this before and reuses a saved skill instead of re-thinking (Section 3).
🧰
Tools
search · browser · vision
The Search toolset: web search, browser automation, vision, image generation, text-to-speech. The core calls whichever the task needs (Section 6).
🧩
Subagents
isolated work
Heavy or parallel work goes to isolated subagents — own conversation, terminal, and Python RPC — so the main context stays clean (Section 4).
🖥️
Backend
hardened sandbox
Code runs on one of five backends — local, Docker, SSH, Singularity, Modal — with container hardening and namespace isolation (Section 6).
Hover a component, or play the flow.
When Hermes runs a subagent's code, where does it execute?

🧠 You've got the Hermes model

One agent, one memory, every surface → six capabilities → memory that auto-saves skills → isolated zero-context-cost subagents → a gateway that runs work unattended → tools (Search) executing across five hardened backends (Experiment). An agent that grows with you.

Learning Reference · Hermes Agent — Nous Research

Get new posts in your inbox

No spam — just new posts and learning pages when they ship.

Share X LinkedIn Reddit Hacker News