tech
Your Agent Picks the Wrong Tool. It’s Not the Agent’s Fault.
Lesson 2.1: When your agent reaches for the wrong tool, the answer isn’t a routing classifier. It’s a description that says when NOT to use the tool.
TL;DR
- Tool selection by AI agents is primarily a description problem, not an architecture or model problem.
- Ambiguous or overlapping tool descriptions lead to unreliable agent behavior.
- Fixing tool descriptions by explicitly stating their limitations and referencing other tools is more effective than adding architectural layers like classifiers or few-shot examples.
- Rewriting descriptions to name competing tools and clarify non-use cases is the most important factor for tool performance.
- Complex architectural fixes often fail at the edges where descriptions still overlap.