Inspiration

Mrs. Chen is 72. Hip replacement. Doctor signed the discharge papers. Medications prescribed. Transport booked. Everything looked fine.

She went home. Two hours later, her daughter called the pharmacy. The medication was not there.

Why? Transport was booked for 5:15 PM. The pharmacy closes at 5:00 PM. Nobody caught it because the transport, the prescription, and the pharmacy hours each lived in a different system.

Software gets tested before it goes live. These plans do not. That is the gap Preflight closes.

What it does

Preflight is an agentic testing system for high-stakes plans. Before a plan executes, it:

  1. Reads the governing policy and compiles it into a checklist of outcome obligations
  2. Generates a live test case in UiPath Test Manager for every obligation, automatically
  3. Runs an adversarial chaos agent that finds hidden cross-system failure cascades
  4. Produces an auditable evidence pack with the exact causal chain
  5. Blocks release if anything critical fails and routes it to a human with a proposed fix
  6. After human approval, re-tests everything, releases, and stores a regression scenario

In the demo, patient PT-1041 looks ready. Preflight finds that transport arrives at 5:15, the pharmacy closes at 5:00, and a required medication can never be collected. Discharge blocked. Nurse approves switching to a 24-hour pharmacy. Re-test passes 9 from 9. Released.

All data is 100% synthetic. Preflight validates operational readiness only. It never diagnoses, prescribes, or changes medication. A human approves every change.

How we built it

  • UiPath Test Manager (Test Cloud) holds 9 severity-tagged test cases generated from policy, all linked to a governing Requirement (PRF:10) for full traceability from policy to test case
  • UiPath Data Fabric is the live system of record for every case, obligation, evidence pack, and human decision. Nothing is hardcoded anywhere
  • UiPath Agent Builder hosts the low-code Obligation Compiler agent that turns a discharge plan and policy into structured obligations
  • Claude Code via UiPath for Coding Agents built the entire rehearsal engine, the chaos search, the evaluator, the remedy proposer, and all UiPath integrations
  • LangGraph powers the adversarial red-team agent that attacks the plan and writes findings back to Data Fabric
  • A live dashboard reads Data Fabric and Test Manager in real time, including contextual help tips served from a HelpContent entity

Challenges we ran into

Cross-system failures are invisible to single-system checks. Modelling the transport plus pharmacy plus medication cascade required building an adversarial search over dependency mutations, not just a checklist.

Embedding rich structured data in Data Fabric, discovering the exact REST endpoints and OAuth scopes for Test Manager, and wiring entity-level permissions each required careful iteration and documentation.

Accomplishments that we're proud of

A working, end-to-end agentic testing pipeline running on UiPath. Policy goes in. Test cases are generated automatically in Test Manager. An adversarial agent finds the failure. Evidence is stored in Data Fabric. A human gates the release. The system learns a regression scenario. All live, all connected.

What we learned

The hardest failures to catch are not inside one system. They happen at the seams between systems. A chaos agent that searches across boundaries finds what a checklist inside a single workflow never will.

We also learned that the "blend of agents" pattern (low-code plus coded plus external framework) is genuinely powerful. Each agent type brings something different, and UiPath is a good place to govern all three together.

What's next for Preflight

Wrap the rehearsal flow in a Maestro BPMN process with a DMN release gate and a real Action Center nurse task. Report test results back into Test Cloud via Orchestrator. Apply the same pattern to insurance claims, payroll runs, and supply-chain releases. Any irreversible action that crosses systems and can harm someone if it fails is a candidate.

Built With

  • 2.0
  • agent
  • builder
  • claude
  • code
  • data
  • fabric
  • langgraph
  • manager
  • oauth
  • python
  • rest-api
  • test
  • uipath
Share this project:

Updates