Suppose a coding agent opens a file, changes a function, then runs a test. The test returns a new failure. The agent now has evidence its first edit did not settle the task. This is an illustrative coding story, not a trace from the ReAct experiments or a claim about one current product.
ReAct joins two kinds of steps. A thought is a written reasoning step in the model's context. An action invokes a tool in an external environment. The tool returns an observation; the model can use that result in its next thought. The paper states that a thought “does not affect the external environment” (printed p. 3). A thought alone cannot supply a missing web page or a new test result.
The paper's own definition, from section 2:
The idea of ReAct is simple: we augment the agent’s action space to  = A ∪ L, where L is the space of language. An action ât ∈ L in the language space, which we will refer to as a thought or a reasoning trace, does not affect the external environment, thus leading to no observation feedback. (printed p. 3)
The authors prompted a frozen PaLM model with human-written example trajectories: growing records of thoughts, actions and observations. The learned model weights stayed unchanged during these prompting experiments. In the paper's question-answering tasks, thoughts appear densely between retrieval steps. In longer interactive tasks, they can appear more sparsely to track a plan or revise a subgoal.
A deliberately small action space
For HotpotQA and FEVER, the paper gave the model three Wikipedia actions:
We design a simple Wikipedia web API with three types of actions to support interactive information retrieval: (1) search[entity], which returns the first 5 sentences from the corresponding entity wiki page if it exists, or else suggests top-5 similar entities from the Wikipedia search engine, (2) lookup[string], which would return the next sentence in the page containing string, simulating Ctrl+F functionality on the browser. (3) finish[answer], which would finish the current task with answer. (printed p. 4)
In short, search[entity] opens a named page or returns similar titles, lookup[string] retrieves the next matching sentence from the current page, and finish[answer] submits an answer. These are the paper's bounded tools, not unrestricted web browsing. The few-shot prompts used six manually composed HotpotQA cases and three FEVER cases.
One FEVER example asks whether Christina Aguilera's Beautiful reached number two on the Billboard Hot 100 in 2003. A lookup for "Billboard Hot 100" on the song's page returns a sentence that supports the chart position but does not give the year. The example's next thought acknowledges that gap, and its finish action chooses NOT ENOUGH INFO. A retrieved result is useful only for the part of a claim it actually supports.
