Building a healthcare agent from intake to submission

A team building an imaging-referral agent needs a shared definition of a completed case. The packet should carry the correct case reference, supporting documents, applicable coverage information, and a summary that links its claims to the source records. It should reach the responsible reviewer with its intended destination visible.

Use an illustrative referral that arrives with clinical records and a missing physician’s order. The first release has a specific responsibility: assemble a review-ready packet, follow up on the missing order, and route the proposed submission for human review. That case becomes the thread connecting the plan in Agent Studio to the tests used before release.

Write the case’s possible continuations

Walking through representative referrals exposes the branches the agent will encounter. A complete packet can proceed to preparation. A missing order needs a request and a later continuation. Conflicting identifiers need evidence and review. For each branch, the team names the expected next state and who owns the next decision.

The referral’s acceptance conditions can be written directly into that working brief:

• The packet and its attachments belong to the selected case.
• The summary identifies the required documents and points to their source records.
• A missing order produces a follow-up that the responsible person can review.
• The proposed submission includes its destination and the evidence behind the action.

Frequency helps the team choose the recurring work for its first release. The range of exceptions determines which branches to include, and the consequence and reversibility of each action help place the review points.

Run the missing-order branch in Agent Studio

The team describes the agent’s role and writes its ordered tasks. Plan Mode presents those tasks as a checklist during execution: identify the case, inspect attachments, retrieve coverage information, prepare the packet, and request review. The team connects the tools for those steps and configures Human Approval where a person’s decision is required.

Workspace holds procedures, source files, and output artifacts. Knowledge scopes make the appropriate policies available. When the agent finds the missing order, it records the gap and prepares a request from the case details. The reviewer receives that proposed action with the packet, destination, policy consulted, and checks already completed.

A correction remains part of the case. If the order carries a different service date, the revised artifact records that change before the next step uses it. When the requested document arrives, the agent can incorporate it into the packet and continue from the recorded state.

Prepared, approved, submitted, and acknowledged each describe a separate event in this procedure. A submission receipt provides evidence that the packet was submitted. The later acknowledgment records the next transition and tells the care team where the referral stands.

Use the same case brief to evaluate the agent

Before release, the team turns the connected procedure into repeatable tasks in Benchmarks. The complete-packet task exercises the standard path. The missing-order task exercises the request and resume steps. The conflicting-record task checks how the agent gathers evidence and asks for review.

Each task checks the expected artifact and final state. Comparing the same cases across changes to instructions, tools, or models reveals which behavior changed. Opening an attempt connects its result to the particular step and records that produced it.

After deployment, operational measures follow those same case groups. Time to a review-ready packet, reviewer corrections, and completion rate each need a defined observation window and data source. Run history and submission events connect the measures to individual referrals.

The next improvement can then target a concrete part of the work: identifying the order, retrieving its source, or giving the reviewer a clearer packet. The case brief remains a common reference for the builders, reviewers, and care team responsible for the outcome.

Explore Agent Studio and Benchmarks.