parallel-probe

Overview ¶

Diagnostic: measure how often the model emits multiple tool calls per assistant turn. ADK already dispatches a single-message multi-call response concurrently (see google.golang.org/adk/internal/llminternal/base_flow.go:585 handleFunctionCalls — sync.WaitGroup over fnCalls). The remaining question is whether the model produces those multi-call responses in the first place. This probe answers that for a real workflow.

Usage (from the repo root, with Vertex env sourced):

go run ./dev/parallel-probe                       # search task, no nudge
go run ./dev/parallel-probe --nudge               # same task, with batching nudge
go run ./dev/parallel-probe --task=multiread      # baseline: 5 independent reads
go run ./dev/parallel-probe --task=multiread --nudge

Output is a per-turn batch histogram so you can eyeball whether the nudge changes behavior before committing to anything.