Crabhaus

What it is

ReAct (Reasoning + Acting), from Yao et al.'s October 2022 paper, has a language model interleave short reasoning steps with actions like searches or tool calls, instead of thinking everything through upfront. Each action's result feeds the next thought, so the model plans, acts, reads what came back, and revises in a loop.

How it works

The model emits a repeating Thought, Action, Observation cycle. A thought states what is needed next; the harness executes the action against an external interface and appends the result to the context before the next thought. Reasoning keeps the plan on track; acting grounds it in information the model does not hold in its weights. The loop ends when the model decides it can answer.

A concrete example

In the paper, ReAct beat imitation and RL baselines on the ALFWorld and WebShop benchmarks by 34 and 10 points in absolute success rate, and Wikipedia lookups curbed the fabricated facts of reasoning-only baselines. A coding agent today runs the same shape: consider the failing test, run it, read the traceback, edit, run again.

Why it matters

Most agent frameworks still run on this loop — a modern tool-calling agent that thinks, invokes a tool, reads the result, and decides again is executing ReAct with structured schemas in place of free-text actions. Its trade-offs are the ones agent builders live with daily: observations curb hallucination, every cycle costs tokens and latency, and the visible trace is what makes agents debuggable.