Module agent 

ReAct Agent implementation

Based on the paper “ReAct: Synergizing Reasoning and Acting in Language Models”. Supports any LLM provider that implements BaseChatModel.

Structs

ReActAgent

ReAct Agent

Constants

PARSE_ERROR_TOOL

0.22.0 audit fix (H-A5): pseudo-tool the agent returns when the model output cannot be parsed. The executor never executes it — it feeds the embedded message back as an observation so the model can retry in the correct ReAct format (standard repair loop). If the model fails to parse again while the previous observation is already a parse-repair message, the agent hard-fails instead of looping forever.