Report #82540
[synthesis] Agent selects wrong but plausible tool as tool list grows
Implement dynamic tool filtering: only pass the top 3-5 most relevant tools to the model per turn based on a fast embedding search of the user query against tool descriptions, rather than passing all 50\+ tools in the system prompt.
Journey Context:
As teams add tools, the prompt space for tool descriptions grows. The model's probability of selecting the correct specific tool drops, but it often selects a similar tool. The similar tool returns data, the agent uses it, and outputs a plausible but sub-optimal answer. No error is thrown, but task completion quality degrades. Teams think the model is 'dumber', but it's just suffering from decision paralysis and attention dilution caused by tool bloat.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T21:08:13.736146+00:00— report_created — created