Tool use fundamentals drill

You are building a calendar assistant for a 2,000-seat company. Users ask things like "move my 3pm with Priya to tomorrow" and "find 30 minutes with the platform team this week". Each model turn costs about 1,500 tokens of tool definitions and the product target is under 8 seconds per request. Design the tool set and defend the boundary of each tool: why it exists, why it is not merged with its neighbour, and how its arguments stop the model making mistakes.

A research agent has a web_search tool. In the trace for one run it called web_search with the identical query "Q3 2026 churn benchmark SaaS mid-market" eleven times in a row, got zero results each time, and the run cost 14 dollars before a user cancelled it. Diagnose the failure and specify what you change in the runtime and in the tool so this cannot recur.

A support agent has a charge_customer tool wrapping your payment provider. One call timed out at 10 seconds, the runtime retried it, and the customer was charged 120 dollars twice. The provider's API took 14 seconds that day because of an incident. Explain exactly why this happened and what you change in the tool, the runtime and the model-facing contract.

A product team wants a data-analysis assistant over a 40-million-row events table in a warehouse. Analysts ask open-ended questions: cohort retention, funnel drop-off, seasonality by region. One engineer proposes twenty purpose-built tools (run_cohort, run_funnel, and so on); another proposes a single Python sandbox with the warehouse client installed. Latency target is 30 seconds per answer and the data includes customer identifiers. Which do you choose, and how do you contain it?