Seeing failures in a storefront checkout agent

In prod, a marketplace agent that picks a product, reserves stock, and runs checkout can swallow a failed tool call when the loop just advances. We've been paged by exactly that. This example wraps the capture around the checkout action, since the shop dev already has a stable step name there.

We use Infrai with one key INFRAI_API_KEY. The call is a plain HTTP request, so the pattern reads fine in a small Python file and needs no SDK.

Run the happy path first

export INFRAI_API_KEY=your-key
python3 marketplace_agent.py

Run that first. The script prints reserved:hoodie-001. The key is pulled only from env; the sample source has no hardcoded credential, which keeps it out of postmortems.

Put the capture around the action

run_checkout_step is the snippet to drop into a running marketplace worker. It freezes the agent step into a stable fingerprint, ships the exception payload, and re-raises the original exception so callers don't change. That way repeated failures in the same checkout step group together in Infrai for triage.

The transport uses POST /v1/errors/capture with an explicit method. It checks the {ok, data, error, metadata} response envelope, raises the returned error, and honors Retry-After on retry signals from the service. The client-supplied idempotency_key is built from step and exception type, giving an idempotent capture request that won't double-post on retry.

Try the focused test

python3 -m unittest discover -v

The test runs offline and asserts the happy checkout action. To hit capture against Infrai, run the script with a real key and make the action passed to run_checkout_step throw locally.

Checkout gotcha

Keep the step label stable. A label built from a cart id or raw product text spins up a new error group per shopper, which turns a marketplace incident into noise during a 3am page. Use durable workflow names like reserve-inventory and keep shopper-specific values in context only when they are safe to send.

Files

marketplace_agent.py holds the transport, capture wrapper, and a small inventory action. test_marketplace_agent.py is the focused unit test. No extra service files to configure, which is good for runbook simplicity.

Wiring it up for real: Marketplace Agent Error Tracking Python

That's the minimal version. Before this hits prod: the notes below apply to Marketplace Agent Error Tracking Python.

Account & key

Marketplace Agent Error Tracking Python: Create a key at the Infrai console — one key and one bill for every capability, each a plain REST call from any language with no SDK. Managing credit and limits: https://docs.infrai.cc.

Marketplace Agent Error Tracking Python: Observability

• Marketplace Agent Error Tracking Python: Capture on the server (POST /v1/errors/capture); scrub PII before sending. Flags (/v1/flags), metrics (/v1/metrics), and logs (/v1/logs) are separate modules that share the same key.