Custom agents that execute recurring workflows

Recurring, well-defined jobs were consuming skilled people's time: work that followed a knowable procedure but required reading context, making judgment calls within bounds, and acting across several tools.

Off-the-shelf automation couldn't hold the whole procedure, and generic AI assistants could draft text but couldn't be trusted to carry a task through to completion.

We treat agents as software systems, not prompts: each workflow is decomposed into steps with explicit inputs, permitted actions, and failure behavior, so the agent's autonomy is a design decision rather than an accident.

Human oversight is part of the architecture. The agents escalate what they shouldn't decide, and their work is logged so it can be reviewed and audited.

Biotite built custom agent systems in TypeScript and Python that execute the client's recurring workflows against their real tools — reading incoming work, taking the bounded actions the procedure calls for, and handing off cleanly when a case falls outside their scope.

The agents run as production services with defined permissions, structured logging, and evaluation against real historical cases before they were given live work.

Recurring workflows that previously required hands-on attention now run through the agent system, with people reviewing exceptions rather than executing every case.

The client retained control over the judgment calls that matter, with a clear record of what the agents did and why.