What we test.

“Get this product page as a shopper in Berlin sees it.”

• Does the agent get the page, location and format it asked for?
• Do the settings it sends actually apply?
• Do errors tell it what to change?

“Index these 200 help articles and return the three that answer this question.”

• Can the agent pick the right model and options from your docs?
• Do rate limits and errors explain what to do next?
• Do results match what the docs promise?

“Transcribe this 20-minute call and label each speaker.”

• Are language and speaker settings applied, or quietly ignored?
• Does the agent know when a long file has finished?
• Do timestamps and formats match the docs?

“Pull the total and every line item from this scanned invoice.”

• Do tables, fields and line items come back complete?
• Does the agent know when a page failed?
• Do output formats match the docs?

“Turn this product photo into a square image on a white background.”

• Do size, format and style settings apply?
• Can the agent tell when a job is still running?
• Are rejected prompts explained?

“Check this new customer against sanctions lists and return the match score.”

• Are scores and thresholds explained well enough to act on?
• Do test and live modes behave as documented?
• Does the agent handle a “needs review” result?

“Forecast next month's demand from this sales history.”

• Can the agent upload data in the format the docs describe?
• Can it track a job through to the result?
• Do results include what it needs to act?