cleandataengineering.aiAgent Harness Simulator

Send a message. Watch the harness work.

You are Rohan, a policyholder of a motor insurer. You have signed in and the claims assistant is open. Type a message, or pick one below, and press Send. The map lights up step by step: every system the message touches, every check that fires, the code that runs, the Python objects it holds, and the JSON that goes to the model and comes back. When the reply arrives, scroll down to see what reached App Insights and Prometheus, and what the evaluations said.

No real agent runs here. The page has a small copy of the agent's routing rules, and it replays what the real claims assistant did on Azure. Steps marked real use replies, timings and payloads recorded from runs on 13 and 14 October 2026. Steps marked illustrative show what the code does, but the values were not recorded from a run.

1 · SendType anything, or click a chip. Ticking a box under "Break something" adds a failure or an attack to the next message.
2 · Follow the numbersEach step has a number. On the map, the badge between two boxes is that step. Bold means the step you are on now.
3 · Open a stepClick a number, on the map or in the list. You see plain words, the code, the variables, and the JSON in and out.
4 · After the replyBelow the steps: the trace, the counters, which alerts would fire, and what the online and release evaluations say.

The map

our harness codea checkgatewaythe modeldata and toolswatch and prove

Click any box to see which steps touched it. Nothing has been sent yet.

Claims assistantsigned in as Rohan · PH-1001
Hello. Your open claims: - CLM-010003: approved - CLM-010004: approved - CLM-010005: registered - CLM-010006: documents pending - CLM-010007: under assessment Your active policies: - POL-010001: KA-01-AB-1234 - POL-010004: KA-03-HK-9090 I cannot say whether something is covered or how much will be paid (an assessor decides), or judge who was at fault. What can I help with?GET /opening · no model called

Everyday messages

The router decides

Attacks in the message

Break something (applies to the next send)

Send a message and its steps appear here, numbered in the order they happen.

The inspector

Nothing selected

Send a message, then click a step to open it. Each step shows:

  • Plain words: what happened, and why it matters.
  • Code: the real file and function that does it.
  • Variables: the Python objects at that moment.
  • Sent and Received: the JSON or HTTP on the wire.
  • The check: which rule ran, and its verdict.

Try "What's the status of my claim?" first. It is the real 7-second turn, with two trips round the loop.

After the reply: what the watchers saw

The trace in App Insights

Counters (Prometheus)

These are the reference agent's counters, which Prometheus scrapes every 15 seconds. On Azure the same numbers go to Azure Monitor as OpenTelemetry metrics, because Managed Prometheus is not built yet.

Alerts on Azure

These are the 6 scheduled KQL alerts in infra/modules/alerts.bicep. Each one runs every 15 minutes.

Evaluations

Find this turn yourself (KQL)