ML//agent//ReAct

ReAct is a control pattern for language-model agents that interleaves reasoning with acting, and it lets an agent adapt to what it finds as it works. The model uses current evidence to select a tool, observes its result, and revises the next decision instead of committing to an entire plan upfront.

ReAct is a control pattern for language-model agents that interleaves reasoning with acting, and it lets an agent adapt to what it finds as it works. The model uses current evidence to select a tool, observes its result, and revises the next decision instead of committing to an entire plan upfront.

Its important contribution, formalizing tool use as reasoning about what to do interleaved with doing it, makes external feedback part of inference rather than an afterthought; the particular prompt format matters less.

Tool errors and surprising observations become opportunities to update the trajectory, provided the harness returns them legibly.

Modern agents often hide or restructure reasoning traces, but the observable control pattern remains.

1Propose2Act3Inspect4Continue