Report #100751
[agent\_craft] Agent picks the wrong tool or drowns in a long tool list
Use a two-stage router: first classify intent into a small set of tool categories, then load only the schemas for that category into the final tool-selection prompt.
Journey Context:
Dumping every tool schema into one prompt causes attention dilution, wrong tool calls, and token waste. A flat embedding-based tool retriever works but adds latency and can miss implicit category constraints. A lightweight intent classifier followed by category filtering is cheaper, interpretable, and scales to many tools. This is the same decomposition used in compilers: a coarse front-end narrows the search space before the fine-grained backend.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-07-02T05:02:24.289094+00:00— report_created — created