You're designing a research agent for a marketing analytics team. You're tempted to give it a powerful tool: web_action, which can browse any URL, extract any content, and submit any form. One tool, maximum flexibility.
Your colleague pushes back, suggesting three narrower tools instead: fetch_article (validates that the URL points to an article and returns clean text), fetch_pricing_page (extracts structured pricing data from a vendor page), and submit_lead_form (only submits forms on whitelisted domains).
Which approach is better, and why?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
Narrow, purpose-specific tools dramatically outperform general-purpose tools in agent systems. They have clearer descriptions ("when do I use this?" is obvious), validated inputs (the agent can't accidentally submit a form on a random domain), structured outputs (pricing data comes back as fields, not raw HTML), and better error messages (failures are scoped). The agent doesn't have to reason about how to use a general capability for a specific purpose — the tool encodes that knowledge. This is a recurring lesson: constrain your tools to known good paths rather than handing the agent a Swiss Army knife.
Flexibility sounds like a virtue and sometimes is, but for production agents it usually trades reliability for surface area. The agent now has to figure out, for every request, how to use web_action correctly — which URL formats work, how to extract structured data from arbitrary HTML, when submitting forms is allowed. Each of these is a place to fail.
It's true that 18 tools is worse than 4–5 for selection accuracy. But three well-differentiated tools is not "too many" — it's the right resolution for a research agent. The selection-degradation effect happens when tools are numerous and overlapping. Three distinct tools are easier to select among than one ambiguous one.
Tool design is one of the highest-leverage decisions in agent architecture. It determines reliability, observability, error handling, and the agent's effective capability. "Either is fine" is the answer of someone who hasn't shipped agents to production.