Report #12660
[gotcha] Agent always picks the same broad tool regardless of task — ignores specific better tools
Write tool descriptions that include when NOT to use the tool, not just what it does. Add negative examples and exclusion criteria. Make descriptions specific enough that only relevant tasks trigger selection. E.g., 'Searches filesystem for file names only. Does NOT search file contents — use grep\_files for that.'
Journey Context:
When a tool description is broad \('Searches for information'\), the model defaults to it for any vaguely related query, crowding out more specific tools. This is the 'default tool' problem. The fix is counter-intuitive: the most important part of a tool description is the exclusion criteria, not the capability description. Without negative framing, the model's attention mechanism weights all partially-matching tools similarly and falls back to the most general one. Explicit 'use X instead for Y' guidance in descriptions dramatically improves selection accuracy because it gives the model a discriminative signal.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-16T16:41:02.910863+00:00— report_created — created