Skip to content

Build a claims-processing workflow, one challenge at a time

Welcome to the ClaimSight Claims MicroHack. You'll build and explore a claims workflow that reads accident statements, retrieves policy evidence, coordinates specialist agents, routes uncertain decisions to human review, and protects sensitive actions.

The workshop is designed to be completed in order. You can choose Bash/zsh or PowerShell in the command tabs; choosing a tab updates matching command tabs throughout the site.

Your path through the workshop

  1. Get your environment ready. Confirm the lab accounts and Azure resources, open the Codespace, and configure your local environment in Challenge 1: Prepare the Environment.
  2. Build the claims workflow. Work through Challenges 2–4 to extract claim details, retrieve policy evidence, and compose the agents into a workflow.
  3. Make it observable and secure. Trace and evaluate decisions in Challenge 5, then test deterministic security controls in Challenge 6.
  4. Understand the design. Review the architecture diagram below before you start.

Before you begin

This is a hands-on lab. You need the assigned lab account, a provisioned Azure environment, and the repository's Claims Hack development container. Start with Challenge 1; it explains the prerequisites and setup.

Choose your first challenge

Challenge What you'll do Time
1. Prepare the environment Verify your lab access, start the Codespace, and configure your tools. 30 min
2. Build the Claims Intake Agent Extract accident statements and find related evidence. 45 min
3. Build the Claims Intelligence Agent Ground coverage decisions in retrieved insurance policies. 30 min
4. Build the Agent Framework workflow Compose the agents and support conditional human review. 45 min
5. Observe and evaluate Trace runs, score decisions, and measure guardrails. 90 min
6. Harden the claims pipeline Defend sensitive actions against prompt injection. 60 min

Challenge 4's Agent Framework workflow is the recommended path. The sequential workflow remains available as an alternative. Challenge 5's optional hosted-agent exercises require a deployment and may incur Azure charges.

Where to run commands

Unless a challenge says otherwise, run commands from the workshop root: the folder containing pyproject.toml, azure.yaml, docs/, challenges/, and agent-framework/. Some tasks explicitly ask you to change into docs/ or agent-framework/; follow the directory shown above each command.

Use the shell selection guide to switch between Bash/zsh and PowerShell, understand which scripts need Bash, and keep secrets out of command output and source control.

Architecture at a glance

The workshop combines specialized Foundry agents with grounded retrieval and deterministic security controls:

flowchart LR CLAIM[Accident statement] --> INTAKE[Claims Intake Agent] INTAKE --> DECISION[Claims Intelligence Agent] DECISION --> REVIEW{Human review needed?} REVIEW -- Yes --> HUMAN[Claims reviewer] REVIEW -- No --> SECURE[Claims Security Action Agent] HUMAN --> SECURE SECURE --> ACTIONS[Protected downstream actions] SEARCH[Foundry IQ evidence] --> INTAKE POLICY[Foundry IQ policies] --> DECISION FIDES[FIDES enforcement] --> SECURE

The workshop instructions are built directly from the challenge files in the repository, so each guide has a single source of truth. Solution walkthroughs remain available in the repository but are intentionally not included in this site.