Why are so many agent tools just 1:1 API wrappers?

I've been spending a lot of time looking at agents that call CRMs, calendars, ticketing systems, and similar APIs. I keep seeing that the model gets blamed for failures that are really caused by the way its tools were designed. Say the agent needs to save a contact in a CRM. If you give it search_contact, create_contact, and update_contact, it has to search first, interpret the result, choose the next tool, and build another request. That looks flexible, but you've pushed a normal if/else branch into the least deterministic part of the system. A single upsert_contact tool is more reliable because the search/create/update logic stays in code. I've started using the same rule elsewhere: Pass info you already have like current user, workspace ID and don't make agent get it by itself every time. Tools return only fields the agent needs, not the provider's entire response object. Validate inputs before they reach the external API (in tool code). Return structured errors instead of a generic 500. Only make tools available when they are relevant to the current workflow. Endpoint-shaped tools still make sense for coding agents or internal tools, where breadth matters and a human is watching. For customer-facing agents taking real actions, I'd rather expose fewer task-shaped tools. Are you using broad, provider-shaped tools in production, or did you end up collapsing them into smaller task-specific tools?