All builds Talk to me
Medical practice · AI agents

Computer‑use agents

Passing practice runs

Two Claude agents that work a medical practice's software through its screen, because a single office can't get an API.

The chart sign-off agent working through a practice queue of charts
The sign-off agent on a practice copy with made-up patients.

How it works

  1. Read the day's list

    Each run starts from the practice software's own report, so the agent works from the same list the staff would.

    Read the day's list
  2. One at a time

    Code hands the model one item at a time in a short, fresh conversation, so a mistake on one chart can't carry into the next.

    The chart sign-off agent working through a practice queue of charts
  3. Code verifies each click

    Code checks every click against what the screen shows. Asked to discard someone else's unsaved note, the agent answers No.

    Code verifies each click
  4. Re-run to confirm

    A second pass confirms the work from the practice app's own record. Routine clicks replay without the AI and stop the moment the screen changes.

    Re-run to confirm

What I built

Payment files agent

It signs in to the practice software over Remote Desktop, exports the day's ERAs (insurance payment files) and hands them to ClaimStride, so payments are read without anyone opening a menu.

Chart sign-off agent

It works the doctor's sign-off queue one chart at a time: it opens the exam, signs it, marks the visit's tests reviewed, and leaves anything unusual on a short list for the doctor.

Guards in code, not the prompt

Code keeps the list, the budget and the record of every step. The agent can't report a click it hasn't seen the result of, and it answers No when asked to discard someone else's note.

Repeat jobs, no AI

Routine clicks it has learned replay with no model at all, so they cost nothing to run. The moment the screen looks different, replay stops and the agent takes over again.

BeforeAfter Payment filesStuck inside the practice softwareExported for ClaimStride by an agentChart sign-offClicked through by hand each weekendExceptions left for the doctor

Engineering

Guards live in code, not the prompt. A report sent in the same turn as a click is refused: it describes a screen the model hasn't seen.

How it fits together
Code keeps the list, the budget and the record. The model gets one short conversation per item, so a mistake can't carry into the next.
Tested
Scored from the practice app's own record. A sloppy script has to fail the scorer.
Runs on
A cloud computer we manage, with no inbound ports.

My role. I built both agents, the practice app they're scored on and the cloud computer they run on.