The AI workforce

Agents are people

Each agent has a row in people with a first name and a job title — Document Control, Reporting, Closeout, and so on — and an agent relationship to the organization. That is what makes their work auditable on the same terms as a human's: the same events, the same attribution, the same access checks. They hold no credentials and claim none; a database constraint forbids it.

Tools, risk, approval

An agent reasons with a language model but acts only by calling a tool from a registry. Each tool carries a risk level — low, medium, high — and each organization's policy can tighten which tools an agent may use. High-risk tools always route to the approval queue: the action is written as a proposal, a person approves, refuses or edits it, and only then does the same handler run. The runtime refuses to unlock high-risk autonomy regardless of policy.

Agents wake on events (a document arrives, a case stalls) and on a daily sweep. Every run records its inputs, the tool calls, their results and the approval that gated them. The activity log shows, per line, who drafted and who decided.

Provider

Gemini today, behind one interface; Anthropic and OpenAI-compatible endpoints are supported by configuration. The provider receives the text of that draft and nothing else.

Removable

The workforce is a non-core module. Drop it and the quality platform runs without it; the shell asks the database whether it is present.

Visible

Anything an agent drafted carries an 'assisted' tag naming the agent and the person who decided.