Preflight — prove high-stakes plans before they touch a person

UiPath AgentHack 2026 · Track 3 — UiPath Test Cloud Branch reality. Break the plan. Prove the outcome. Then act.

Preflight forks a live case into a disposable digital twin, compiles the case's promises into executable outcome obligations, unleashes adversarial agents to break every dependency, proves the outcome through UiPath Test Cloud, and lets the real action proceed only after the plan survives rehearsal — or a human accepts a documented exception in Action Center, with UiPath Maestro as the governing release gate.

Demo domain: safe hospital discharge (100% synthetic data). Preflight validates operational readiness only — it never diagnoses, prescribes, or changes medication, and a clinician stays in control.

---

The business problem

Enterprises automate consequential, often irreversible actions — discharging a patient, settling a claim, running payroll, releasing a shipment. Today these execute optimistically: the workflow assumes every downstream dependency will hold. When one silently fails — the pharmacy is closed when transport arrives, a medication isn't covered, a follow-up was "recommended" but never booked — the failure is discovered after it has already harmed someone. Software is tested before production; human-impacting plans are not. Preflight closes that gap.

What makes it different

• vs. agent-evaluation tools (Giskard/LangSmith/Foundry): those test the agent/model. Preflight tests the planned real-world outcome of a specific case and gates the real action.
• vs. UiPath Process Simulation: that predicts aggregate efficiency from historical data. Preflight proves this one live case's safety/outcome obligations and produces an executable release gate.

How it works (the pattern)

Live case → FORK twin → COMPILE obligations → CHAOS attack → PROVE (Test Cloud)
          → [GATE] ──fail→ Action Center (human remedy) → re-test ↺
                    └─pass→ Release for real execution → LEARN (regression scenario)

UiPath components used

Component	Role
Test Cloud / Test Manager	Core engine — per-case test set, execution, results, traceability (the gate's proof).
Maestro (BPMN + DMN)	Orchestrates the rehearsal; DMN is the hard release gate; agents = Service Tasks, human = User Task.
Agent Builder	Low-code agents: the Obligation Compiler (Context-Grounded on policy) and Remedy Proposer.
Coded agent (+ optional CrewAI/LangChain governed by UiPath)	The chaos / red-team engine that searches failure combinations.
Action Center	Human-in-the-loop: nurse approves/edits a remedy or accepts a documented exception.
Data Service	Source of truth for cases, obligations, results, decisions, evidence, regression library.
API Workflows	Synthetic pharmacy / insurer / transport / scheduling / notification adapters.
UiPath Apps	The frontend — every value bound to Data Service (no hardcoded data) + an [i] help/tour system on every screen.
UiPath for Coding Agents (Claude Code)	Used to build the engine — see docs/CODING_AGENT_EVIDENCE.md.

Agent types used: BOTH — low-code (Agent Builder) and coded agents, plus an optional external framework (CrewAI/LangChain) governed by UiPath, plus a coding agent (Claude Code) used to build it.

Full platform mapping: docs/UIPATH_MAPPING.md.

---

This repository

This repo contains the platform-independent reference engine for Preflight: the exact logic (compiler, chaos/red-team search, obligation evaluation, gate, remedy, evidence) that is ported onto the UiPath platform. It runs with zero third-party dependencies so judges can verify the core behavior in seconds, then watch it run on UiPath in the demo video.

preflight/
├── engine/            # the reference engine (pure standard library)
│   ├── models.py      # entities = Data Service schema
│   ├── adapters.py    # stub services = API Workflows
│   ├── compiler.py    # case-to-test compiler = Agent Builder agent
│   ├── chaos.py       # chaos/red-team agent = coded agent
│   ├── evaluator.py   # obligation execution = Test Cloud
│   ├── remedy.py      # remedy proposer = Agent Builder agent
│   ├── evidence.py    # evidence pack
│   └── orchestrator.py# the flow = Maestro process
├── data/
│   ├── cases/         # synthetic discharge cases (golden, clean, coverage)
│   ├── policy/        # discharge readiness policy (compiler grounding)
│   └── help/          # [i] help + tour content (backend-served)
├── scripts/run_demo.py
├── tests/test_end_to_end.py
└── docs/              # UiPath mapping + coding-agent evidence

Prerequisites

• Python 3.10+ (developed on 3.12). No pip install needed for the reference engine.

Run it

cd preflight

# Full narrated rehearsal of the golden case (blocks -> remedy -> re-test -> release)
python scripts/run_demo.py golden_case

# Other cases
python scripts/run_demo.py case_clean      # passes first time (green path)
python scripts/run_demo.py case_coverage   # multiple failures, then recovers

# Tests
python tests/test_end_to_end.py            # or: python -m pytest

Expected: golden_case is BLOCKED on the critical "medication collectable" obligation (transport arrives 17:15, pharmacy closes 17:00), the nurse approves a 24h-pharmacy remedy, the re-test passes at 100%, the plan is RELEASED, and an evidence pack is written to out/.

Setup on UiPath Automation Cloud

See docs/UIPATH_MAPPING.md for the component-by-component build/run guide (Data Service entities, Agent Builder agents, Test Cloud test sets, Maestro process, Action Center app, UiPath Apps frontend).

Safety & data

• 100% synthetic data; no real PHI/PII anywhere.
• Preflight validates operational readiness only; it does not diagnose, prescribe, or change medication. Every clinical decision stays with a human.

License

MIT — see LICENSE.