A finance ops agent has 22 tools: lookup_customer, lookup_account, lookup_invoice, lookup_payment, lookup_subscription, lookup_tax_record, lookup_credit_memo, and 15 others. Tool descriptions are well-written. Tool selection accuracy is still 64%.
What's the most effective fix?
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.
Tool selection accuracy degrades as tool count grows — typically around 12+ tools, even with excellent descriptions. The fix is structural: give the primary agent 4–7 high-frequency tools and route niche work to specialized subagents with narrower, role-appropriate tool sets. The primary agent's decision space stays small; specialized work happens where the tool set is purpose-fit.
Description quality matters, but it has diminishing returns. Once descriptions are clear, additional detail just adds tokens. The primary lever above 12 tools isn't description detail — it's decomposition.
tool_choice: "auto" is already the default and doesn't change the fundamental problem of selecting among 22 similar tools. This is reaching for a configuration knob without diagnosing the actual cause.
Stronger models do better on tool selection at the margin, but they don't close the structural gap. A 22-tool agent will still underperform a 5-tool agent with 5-tool subagents — and you've also paid more per call.