Home / Agentic systems
Agentic AI systems that do multi-step work under your rules
An agent is a model that can take several steps toward a goal, using tools you give it. They use an open-weight model: the untuned 4B on our hardware, or a larger one through OpenRouter. If you turn training on, a model of your own, run on our hardware, takes over once it is trained. The agents running today are in our founder's own business, HousingInteractive, and those currently use cloud models.
What it is
A chat model answers a question. An agent carries out a task: it reads a ticket, looks up the customer, checks a document, drafts a reply and hands it to a person. We design each agent around one workflow, with a fixed set of tools and clear limits on what it may do.
Once it runs on a model trained for you on our hardware, its behaviour does not change when an outside provider updates or retires a model.
Who it is for
Teams with a repeated process that involves reading, deciding and writing across several systems, such as support triage, internal research, document review or onboarding, and that want it done under their own rules.
What you get
- A working agent. Built for one workflow, connected to the systems it needs and nothing more.
- Scoped permissions. A written list of what the agent can read, change and send, enforced in the web app and the gateway and not left to the model. Enforcing it inside the hosted desktop is in progress.
- Human approval steps. Actions you choose to review, such as sending an email or changing a record, wait for a person.
- An audit trail. Every step, tool call and result is logged in a form your staff can read.
How it is done
- Map the workflow. We follow the process as people do it today and mark which steps are safe to automate.
- Choose the model. An open-weight model to start: by default a larger one through OpenRouter, or the untuned 4B on our hardware. Training one for the agent's steps on our hardware is in development.
- Build and test. We run the agent on past cases and compare its work with what your staff did.
- Roll out in stages. It starts by suggesting, then acts with approval, and only acts alone where you decide it should.
What we need from you
- One workflow, and someone who knows how it is done today.
- Past cases we can test against.
- Access to the systems the agent will use, on the limits you set.
- Decisions on which actions need a person to approve.
Questions, answered.
Can an agent act without anyone checking it?
Only where you decide it can. The default is that it suggests and a person approves. Each action type has its own setting.
What stops an agent from doing something it should not?
In the web app and the gateway, permissions are enforced by the surrounding software, not by asking the model to behave. Enforcing them inside the hosted desktop itself is in progress. If the agent has no access to a system, it cannot use it.
Does the agent need a large model?
Often not. Many agent steps are narrow, and a small model can handle them. We test this on your cases before choosing.
Other services
- DistillationTrain a compact model for one task of yours.
- InstallationOn your own servers. Scoped per client, not yet delivered.
- HostingYour workspace and desktop on hardware we operate.
Start on a practice company.
Request an invite and try it on made-up customers first. Nothing reaches real people while you learn.