What it runs on

Your IT will ask what this is built on, so here it is in plain terms.

The models

Agents run on Anthropic’s Claude models, under a business agreement: no training on your data, and retention only as those terms allow. Anthropic is our only model provider, which means one set of terms in your file and no question about where a document went.

Most of what happens never reaches a model at all. A step that runs as a script sends nothing, and most steps are scripts. See Script first, agent when needed.

The runtime

The part that gives an agent its tools, its turns and its access is Anthropic’s Claude Agent SDK, running inside the container that belongs to your company. It is not a shared machine and not a browser session pretending to be a colleague.

We did not build our own agent harness, on purpose. Keeping one correct is a full time job for a company whose whole product is that harness, and the companies that build their own end up maintaining it instead of using it. You do not build one either, which is the point of paying for this at all.

The part that is ours

The harness makes an agent possible. It does not make it something you can hand a process to. That is the layer we build:

• A plan approved before anything is built. See Jobs and plans.
• A gate before it goes live. Scored on your own cases; below the agreed number the switch stays off. See Measured before it goes live.
• Isolation. One container per customer, with only the connections that job needs. See Where your data lives.
• Script first. The mechanical work costs nothing and gives the same result every time.
• Written norms and rules, so behaviour does not depend on how a question was phrased. See Norms and rules.
• Three ways out, decided in the plan rather than by the agent at runtime. See Three ways out.
• A timeline with a cost on every run, and an ROI report that says when a job stops being worth it.

What this means for you

Model versions change over time. Keeping up with them is our work, not yours, and a change that affects a job is measured on your own cases before it is used, the same way the job was measured the first time.

Your data is not training data. Your records stay in your systems. What goes to a model is the document or record a step is working on, and nothing else.