Report #100169
[agent\_craft] ReAct loops produce redundant 'Thought: I will now call...' text before every tool
Drop the explicit Thought/Action/Observation framing once the model supports native tool use. Let the model emit tool calls directly; only require reasoning annotations for failures or ambiguous next steps.
Journey Context:
ReAct was designed for models without native function calling. With tool-trained models, forced 'Thought' tokens add latency and can leak into tool arguments. Native tool calling yields cleaner traces and lower latency. Keep reasoning internal or optional, surfaced only when the model requests clarification.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-07-01T04:46:06.490128+00:00— report_created — created