Tool calling

• The model must fetch or mutate real data (search, DB, CRM, calendar).
• You need precise outputs or actions, not just text.
• You want better observability and control than free-form prompts.

Tool calling shifts reliability from prompt-only to contract-based execution. PMs define which actions are safe, what to log, and how to approve or sandbox risky calls. It impacts latency (network) and cost (extra calls) but boosts determinism and safety.

Start with read-only tools, add write tools behind approvals. Provide strict schemas, defaults, and examples. Monitor success rate, retries, and cost per tool. In 2026, add automatic fallbacks (e.g., clarify with user, choose alternative tool) and enforce per-session budgets.

A support bot calls a ‘get_subscription’ tool before suggesting fixes. Tool success climbs to 97%; incorrect plan recommendations drop 30%, and average handle time improves by 18%.

• Defining vague tool schemas that accept anything, causing partial failures.
• Letting the model loop on failing tools without limits.
• Not logging arguments, making debugging impossible.