Three layers, one agent
The modeldecides what to do
→
The harnesstools, memory, skills, schedules, checks
→
AgentSmithyour skills, your systems, run every day
The model reads what you ask and decides what should happen next. The harness is the software around it that
actually does things, like reading files, searching the web, remembering you and running jobs on a schedule. The
third layer is ours. We install skills for your kind of work, connect the agent to the tools your team uses, give it
its own private environment and look after it every day.
The model brings the intelligence, the harness makes it dependable, and we make it yours.
Most AI products give you only the first layer and a chat window. That's why they impress in a demo and then let you down on real work.
The model decides, it doesn't do
A language model is very good at reading a messy request, working out what you mean and planning the steps. It is
bad at exact facts, because it writes the most likely answer, not a checked one. Ask a model on its own what a client
owes you and it gives you a confident number, and if you ask again you might get a different one.
So in your agent the model never gets to be the source of truth for a number, a price or a date. It decides which
tool to use, and the tool fetches or calculates the real value.
You also don't pick or manage models. Every request your agent makes goes through our own gateway, with a key that
belongs to your agent alone. We route each kind of work to a model that suits it, for example text to one model and
images to a vision model, and we change them when something better comes out.
The harness turns words into actions
Your agent runs on Hermes Agent, an open-source agent framework from Nous Research. It's the layer
between the model and the real world, and it gives your agent these abilities.
- Real tools. It can work with files, run commands and code, search the web and read pages, and reply with voice. Every action goes through a defined tool with checked inputs, so the model can't invent an option that doesn't exist.
- Memory. It keeps a short, curated memory of you and your business and loads it at the start of every conversation.
- Searchable history. Every conversation is stored, so it can look up what you decided last month instead of asking you again.
- Skills. These are written procedures it loads only when a task needs them. It can also write new skills from work it has done, so a routine it has done once becomes a routine it does the same way every time.
- Scheduled jobs. Say "every Monday at 9am, send me last week's enquiries" and it runs on its own and posts the result back to your chat.
- Helpers in parallel. For a bigger job it can hand pieces to up to ten helper agents at once and combine what they find.
- Guardrails. A hard block on destructive commands always applies, secrets are removed from its output and logs, and files that hold passwords and keys are off limits.
Here's what that looks like when you ask a question with a number in the answer.
# you ask: "How much does Acme owe us?"
search_files accounts/acme/invoices ok 6 open invoices
execute_code sum(open invoices) ok 47,380.00
check total against the ledger ok matches
# reply: "Acme owes $47,380.00 across 6 open invoices. I checked the total against the ledger."
# if the check had failed, it would stop and tell you instead of guessing
Illustrative figures. The tools and checks for your business are set up during the build.
What we add on top
Hermes Agent is open source, so anyone with the skills can install it on a server. What you get from us is
everything between a fresh install and an agent you can trust with part of your business.
New skills early, like BrowserSkill
Most AI products only get a new ability when the vendor decides to ship it. Your agent is open and we configure it,
so we can add a new capability as soon as it has proved itself. BrowserSkill is a good example. It's an open-source
browser extension, made by Tencent, that lets your agent work inside your own Chrome or Edge, on sites where you're
already signed in. It can open pages, read them, fill in forms and click through portals that have no API.
We run the server side for your agent, give you a one-time pairing link in your Command Center and set clients in
mainland China up with Edge. Your passwords stay in your browser, and the agent stops and hands back to you for logins,
captchas and payments. The big assistants have started to add browser control too, but few let you send an always-on
agent a message from your phone and have it do the work in the browser on your office computer.
See how browser access works.
Skills built for your work
We write skills for the jobs your business repeats. For a property firm we built one that turns architectural
drawings into a measured bill of quantities, and another that runs feasibility numbers in code so the model never does
the arithmetic. Your agent also writes its own skills as it learns your routines, and we keep and review them.
A knowledge base from day one
We set up a wiki in your agent's workspace, a folder of linked notes with an index, and we instruct the agent to
check it first and keep it up to date. It's where your procedures, price lists and client history live.
How to use it.
Answers it has to check
When we build a workflow, we decide where each fact comes from. Prices, balances and availability are read from
your records and calculated in code, and anything a check can't confirm goes to a person.
Your own private environment
Each client's agent runs in its own isolated container on its own network. It can't see other clients' agents or
our internal systems, and it has fixed limits on memory and processing. We connect your accounts for you, so you never
paste a password or an API key into a chat.
Tested before you get it, watched after
An agent isn't handed over until a live check passes. That covers the model and its fallback, voice notes, images,
long-conversation handling and who is allowed to message it. After that, a health check runs across every client agent
each night. Updates are pinned to a tested version and moved one agent at a time, with a way back. When an update from
the open-source project broke one of our models, we fixed it across every client agent before the official fix came out.
Costs kept in check
Because every request goes through our gateway, we see unusual spending as it happens. An alert fires when an agent
uses three times its normal amount, and most of each request is served from a cache, which keeps the running cost down.
Continuity
Your agent's memory, wiki, skills and conversation history are stored separately from the software, so restarts and
updates don't touch them.
Side by side
| A chat app | An agent you install yourself | Your AgentSmith |
|---|
| Works while you're away | No, it only answers | Yes, once you've set up a server | Yes, hosted and running all the time |
| Connected to your systems | Only the vendor's integrations | If you build the connections | We connect the ones your team uses |
| Skills for your work | None of your own | You write them | We build them, and it writes more |
| Works in your signed-in browser | Depends on the product | If you install and secure it | Set up and paired for you |
| Privacy and isolation | A shared platform | Your responsibility | Its own environment per client |
| Updates | Whenever the vendor ships | Your job, and they can break things | Tested, then moved one agent at a time |
| Who to call | A help centre | Nobody | Us |
What it looks like in an ordinary week
| Situation | What your agent does |
|---|
| Monday, 9am | A scheduled job pulls last week's enquiries and sales from your systems and posts the summary in your chat before you open your laptop. |
| A customer writes at 11pm in Chinese | It answers in Chinese from your price list and availability in the wiki, and leaves anything it can't confirm for your staff in the morning. |
| A supplier portal with no API | You message it from your phone, and it logs the order through BrowserSkill in your office browser, stopping for you at the payment step. |
| Month end | It totals open invoices in code, checks the figure against the ledger and flags the two that don't match. |
| A task you've explained twice | It saves the steps as a skill, so the third time you only say the skill's name. |
Why this is the part that matters
Nearly every business that tries AI hits the same wall. The demo is impressive, then you point it at real work and
it makes up a number, forgets what you told it last week or simply isn't running when you aren't typing.
None of what's in this guide makes the model smarter. The harness makes it dependable, and the work we do around it
makes it fit your business and keeps it running, so a weekly report arrives every Monday rather than four Mondays out of five.