06

Tool Selector

jev primitiveChoice per step (multi-step, traced)
labelsget_weathersearch_webcalculatordatabasecreate_ticketsend_emaildatabase > calculatornoneask_user
dataset20 labeled rows
tracingLangfuse

An agent that loops: pick the next tool (or stop), write its arguments, run it, look at the result. Jev picks from a closed list (6 tools, plus finish and ask_user) and reports a confidence. An LLM writes the arguments, and only for the chosen tool. The tools are offline fakes, and every loop stops at 4 tool calls. The label is the whole trajectory (e.g. database > calculator). Traced: every loop turn and every model call, with its cost, in Langfuse.

With Jev

◐ ▮▮▮ ▦

Without Jev

LLM · CODE
Examples

Calls Jev and every baseline for real, traced in Langfuse. Nothing is saved.