We run a good part of our company on AI agents now. Not a chat widget on the website, but actual agents that pick up work, remember what they did, and act on a schedule without anyone prompting them. We build them on Hermes, the open agent framework from Nous Research, and we run the models behind them on hardware we own (we wrote about that setup in Running Enterprise AI On-Prem). Over the past few months this went from a weekend experiment to something we lean on every day, both for the business and at home. This post is about how we provision these agents, why we lock them down the way we do, and why we can build the same thing for your company.
A model is not an agent
A model answers a question and then waits for the next one. An agent has a job. It has its own memory, a defined set of tools, and the ability to do things on its own. Hermes is the framework we use to define those agents, give each one a role, and keep it inside boundaries we set. The model is the engine. The agent is the employee.
One agent per job, and they stay in their lane
We do not run a single assistant that tries to do everything. We provision a separate agent for each area of work: running the business, client and development projects, research, personal tasks, and keeping the household organized. Each agent has its own memory and its own short list of tools, and it can only see and touch what we have granted it. The agent that handles client work has no path into personal accounts. The one that runs the family calendar cannot reach anything to do with the business. That separation is the entire point. It is what makes it safe to hand an agent something sensitive.
Reaching them without yet another app
We talk to our agents through Signal, the same encrypted messenger we already use. Behind it we built a small router so one Signal number can fan out to every agent, each on its own private channel. You message an agent the way you would text a coworker. Provisioning a new one comes down to standing up its profile, wiring its channel, and deciding who is allowed to talk to it. There is no new dashboard and no portal to babysit.
We start with security, not bolt it on later
These agents can run commands, edit files, and call APIs, so we treat their access the way we treat production infrastructure. The models run on our own hardware, tied into our local inference stack, which means prompts and client data never leave our environment. Every agent runs sandboxed with a short, explicit list of what it may use, and only approved people can reach it. Nothing gets a capability it does not need for its job. When a client cares about where their data lives, this is the answer.
Putting them to work on engineering
The change we feel most is that our agents drive our own development tools. An agent can take a task, run it through Claude Code or another coding harness, make the change in a sandbox, check it, and report back, all from a message. Dependency updates, migrations, and routine maintenance can run on a timer with nobody at the keyboard. Because each agent keeps a record of what it has done, the setup gets more useful the longer it runs rather than less.
The same system runs our house
The family agent keeps everyone’s calendars straight, sorts out scheduling, and tracks what needs to happen and when. It sounds small next to the engineering work, but it is the clearest proof that this is dependable and not a demo. We trust it with the real logistics of our week.
What this means for your company
Here is why we bothered to write this down. Standing up secure agents is mostly an integration problem, not a model problem. It takes infrastructure you can trust, tight access control, a sane way to talk to the agents, and real connections into the systems a business already runs on. That is the work we have been doing for more than fifteen years: building web platforms and the APIs that hold them together. Agents succeed or fail on exactly those connections, into calendars, codebases, internal tools, and third party services.
We can provision the same kind of system for you. Custom agents built on Hermes, scoped to your data and your rules, running on your hardware or a secure cloud you control instead of someone else’s. We built it for ourselves first, so we already know where the sharp edges are. If you want agents your team can actually trust with real work, get in touch.