ReAct: Why an Agent Needs a Loop, Not Just a Plan
ReAct is a prompting paradigm for language-model agents that interleaves verbal reasoning with actions in an external environment. The important question is not simply whether a model can produce a plausible explanation or call a tool. It is whether reasoning can guide the next action, and whether the result of that action can change what the model believes it should do next. This series uses ReAct to build a practical mental model of that loop: what it adds to language-model task solving, why it helps in some settings, and where its gains give way to search errors, repeated steps, or unsafe control.
That focus matters because reasoning and acting solve different parts of an uncertain task. A model may know how to write a convincing chain of sentences yet have no way to check whether the facts are current. It may also know how to emit an action such as a search or a navigation command while failing to understand what the result means for the larger goal. ReAct is the paper's attempt to join these capabilities without treating either one as sufficient on its own.
The gap between reasoning and action
Reasoning-only prompting gives a language model room to derive an answer from the information already in its context. Chain-of-thought prompting made this approach especially visible: instead of producing only a conclusion, the model writes intermediate steps that can organize arithmetic, comparisons, or multi-part questions. But those steps remain inside the prompt. Unless the model is given a separate source of feedback, a mistaken fact introduced early can support several later steps. The resulting answer may sound coherent precisely because the unsupported premise has been carried forward consistently.
Action-only behavior has the opposite weakness. An action can reach an external source, move through a simulated room, or select an item from a product page, but an action by itself does not explain the task state. In the paper's examples, an action-only agent can search a household location, fail to find the desired object, and continue producing related actions without recognizing that its previous assumption was wrong. The problem is not that the agent has no observations. It is that nothing in the trajectory necessarily turns an observation into an explicit plan update.
ReAct is designed around the gap between these two failures. In a knowledge question, the model may need to identify an intermediate entity, retrieve a fact about it, and then use that fact to choose the next search. In a household task, it may need to break a goal into subgoals, remember which one is complete, and decide where an object is likely to be found. In both cases, the model must connect what it thinks with what it does, then connect what happened with what it thinks next.
This can also be understood as a question of state ownership. In reasoning-only prompting, the model owns the working story but has no built-in mechanism for updating that story from the world. In action-only prompting, the environment can return new events, but the trajectory may not contain a durable interpretation of what those events mean for the goal. ReAct puts both forms of information into one context: the model proposes an interpretation in language, the environment confirms or contradicts it through an action result, and the next thought has access to both. That does not make the context correct by itself. It makes the source of a correction visible enough for the model, and for a human inspecting the trajectory, to act on it.
The distinction is especially important when an action fails. A missing search result is not the same as evidence that an entity does not exist; a command that produces no change is not the same as completing a subgoal; and a product that matches one attribute is not necessarily the product that satisfies the whole instruction. ReAct is useful only if its reasoning traces preserve these distinctions and use them to select a better next action. The loop therefore turns feedback into a design responsibility: someone—or some controller—must still decide what counts as informative feedback and when the agent should stop trying.
The contrast is useful, but it should not be mistaken for a claim that one mode always wins. ReAct is a way to make the relationship between internal reasoning and external feedback explicit enough to test. The paper evaluates that relationship across information retrieval, fact verification, household interaction, and web navigation, using reasoning-only and action-only variants as controlled comparisons.
What the ReAct loop changes
The paper describes ReAct as an interleaved trajectory of thoughts, actions, and observations. A thought is a piece of language that helps the model organize the current context. It can decompose a goal, extract an important detail from an observation, suggest a likely location, notice an exception, or mark a subgoal as complete. The thought itself does not change the external environment. An action does: it searches, looks something up, navigates, selects, or otherwise interacts with the task environment. The environment then returns an observation, which becomes part of the context available for the next decision.
This creates two directions of influence. Reasoning is used to decide what to do—what entity to search, which location to check, or whether a candidate item satisfies the user's constraints. Acting is used to improve reasoning—by returning a fact, a failed search, a changed state, or evidence that contradicts the previous plan. The model is not merely writing a rationale before acting. It is repeatedly using language to shape the next probe and using the probe's result to reshape the language that follows.
Consider a multi-hop question. A model might first reason that it needs to find a show, then identify the hotel where the show is performed, and finally retrieve the hotel's number of rooms. A search result can reveal that the first query was too broad; the model can reformulate it rather than treating the empty result as an answer. The value of the thought is therefore practical: it directs information gathering and records why the next action is relevant. The value of the observation is equally practical: it prevents the model from having to rely only on what was present before the search.
The same structure applies when the environment is a text-based household. A goal such as placing a clean object on a surface contains several dependent actions: find the object, take it, clean it, and place it. A useful thought can keep those subgoals in order and distinguish “the object has been found” from “the object is ready to place.” The environment remains the authority about what actually happened. If an action produces “nothing happens,” the trajectory has a chance to treat that result as a state change in its plan rather than silently repeating the command.
The loop does not require a thought before every action. In the paper's knowledge tasks, thoughts and actions are interleaved densely because each retrieval result may determine the next reasoning step. In longer decision tasks, thoughts are deliberately sparse and can appear where planning, progress tracking, or exception handling is most useful. This matters because narration is not the objective. The objective is to give the agent an intermediate language channel that helps it control a changing task without confusing language about the task with a change in the task itself.
Why the paper tests more than one kind of task
ReAct's subject is a general interaction pattern, but the evidence comes from specific benchmark environments. HotpotQA tests multi-hop question answering over Wikipedia passages. FEVER tests whether a claim is supported, refuted, or left unresolved. In the paper's question-only setup, the model does not receive the supporting paragraphs in advance. It can rely on internal knowledge or use a deliberately simple Wikipedia interface with search, lookup, and finish actions. That interface is weaker than a state-of-the-art retriever, which makes the model's information-seeking choices visible.
The other tests change the shape of the problem. ALFWorld presents household-like environments in which a high-level goal may require more than 50 actions and many possible locations. WebShop presents noisy product descriptions and user instructions that combine attributes such as size, color, scent, and price. These settings ask whether the same thought–action–observation relationship can support a long sequence of decisions, not just a short answer.
These environments are complementary rather than interchangeable. HotpotQA and FEVER make the information loop easy to isolate: the central question is whether the model can seek and use the right evidence before it commits to an answer or label. ALFWorld makes persistence and state tracking visible: the model has to keep a high-level goal intact while individual actions change what can be found next. WebShop adds a different kind of uncertainty, because the model must connect a user's several constraints to incomplete and noisy product observations. Seeing ReAct in all three shapes prevents a narrow interpretation of the method. It is not merely a search wrapper, a household planner, or a shopping policy; it is an attempt to use the same language-mediated feedback relationship under different pressures.
The results already suggest a conditional picture. On HotpotQA and FEVER, ReAct improves on action-only prompting, reaching 27.4 exact match and 60.9 accuracy compared with 25.7 and 58.9. It does not simply dominate reasoning-only prompting: chain-of-thought scores 29.4 on HotpotQA but 56.3 on FEVER. On the interactive tasks, the reported gap is more pronounced. The best ReAct trial reaches 71% average success in ALFWorld, compared with 45% for the best action-only trial, while WebShop reports 40.0% success for ReAct versus 30.1% for action-only prompting. These figures describe the paper's models, prompts, environments, and metrics; they are evidence for a mechanism under test, not a universal ranking of agent designs.
The paper also reports practical reasons to inspect the whole trajectory rather than just the final score. External retrieval can reduce unsupported facts, but an uninformative search can derail the next step. Interleaving can make a decision easier to inspect because a reader can see which information came from the environment and how the model connected it to an action. Yet a visible thought is not automatically a faithful record of every hidden computation. ReAct makes reasoning and acting more legible as an operational sequence; it does not remove the need to test whether the sequence is correct.
That is the central tension this series will carry forward. ReAct does not add value merely by producing more text or by attaching a tool to a language model. Its distinctive claim is that language can serve as a control state between environment interactions: it can organize a plan, target evidence, and record a revision. The next question is therefore precise. What, exactly, happens inside that interface when a thought is treated as part of the agent's state, an action crosses into the environment, and an observation returns to change the plan?