The ReAct Pattern: Thought, Action, Observation

The ReAct pattern (Reasoning and Acting) is an agent design pattern in which an LLM alternates between reasoning (Thought), calling a tool (Action) and reading the result back (Observation). This loop repeats until the agent produces a final answer. It was introduced by Yao et al. (arXiv:2210.03629, October 2022, ICLR 2023). The reasoning steers the choice of tool, and the tool observations correct the reasoning.

• What it is: A loop of Thought, Action and Observation in which an LLM interleaves thinking and acting and uses tools within the same context.
• What it is good for: Tool-supported, factually grounded tasks with a fully traceable history. ReAct is the standard starting point for agent projects.
• What to watch out for: Infinite loops, reasoning drift and hallucinated tool arguments; mitigated with hard iteration limits and function calling.

Where ReAct comes from and which problem it solves

ReAct emerged as a response to two opposing weaknesses of earlier approaches. Pure Chain-of-Thought (CoT) reasons exclusively internally and therefore hallucinates facts, because it lacks any grounding in the outside world. Pure action agents, such as early WebGPT or SayCan approaches, conversely cannot reason abstractly about long-term goals or recover from exceptions.

ReAct unites both: the verbal reasoning sits in the same context as the actions. This gives the LLM a working memory across the entire trajectory and makes the agent interpretable. Formally, ReAct expands the action space: alongside external actions that trigger real observations, there are linguistic "actions" (the Thoughts), which change nothing in the environment but only update the agent's internal context. Notably, the pattern already works with one to two in-context examples, without fine-tuning.

The loop in detail: Thought, Action, Observation

The canonical sequence is: User Query → [LLM: Thought] → [LLM: Action] → [Environment: Observation] → [LLM: Thought] → … → Finish[Answer]. The three steps interlock as follows.

This cycle runs until the model emits Action: Finish[Answer]. The practical upper bound is usually around 10-25 steps before context loss or reasoning drift take over.

Pseudocode of a ReAct loop

1. context = systemPrompt + toolDescriptions + userQuery
2. step = 0
3. MAX_STEPS = 15 # hard limit against infinite loops
4. while step < MAX_STEPS:
5.     output = LLM(context) # produces Thought + Action
6.     if output.action == "Finish":
7.         return output.action.argument # final answer
8.     observation = runTool(output.action) # external call
9.     context = context + output.thought + output.action + observation # append observation
10.     step += 1
11. return "Limit reached, no solution found." # clean abort

The central point: in every iteration, the entire history so far is sent to the LLM again. This is the source of both the strength (complete memory) and the most important weakness (cost).

Strengths and limitations from the research

The original paper documents ReAct's advantages across several benchmarks. Importantly, the figures stem from the conditions of 2022/2023 (GPT-3 and PaLM class) and should be read as relative values, not as today's absolute values.

• ALFWorld (text-based household tasks): +34 absolute percentage points over imitation/RL baselines, with one to two examples.
• WebShop (e-commerce navigation): +10 absolute percentage points in success rate over IL/IL+RL.
• HotpotQA and Fever: competitive with or better than pure CoT or pure action generation; the strongest configuration is a hybrid that switches between internal knowledge and tool-supported reasoning.

On the cost side, ReAct is expensive and slow. The token cost grows with O(N · T): N steps times the cumulative context length, because every step pays for the prefix (system prompt, tool descriptions, entire history) again. The latency is strictly sequential, i.e. roughly N × (LLM response time + tool response time). It is precisely this prefix repetition that is the weakness addressed by the successor pattern ReWOO through just two LLM calls.

Failure modes and countermeasures

Three failure modes occur particularly frequently in practice:

• Infinite loops: Without a hard step limit, the agent can reason in circles indefinitely. Countermeasure: explicit limits such as recursion_limit in LangGraph, max_tool_iterations on the AssistantAgent in AutoGen (alternatively a termination condition such as MaxMessageTermination), max_iter in CrewAI. LangGraph's create_react_agent offers a remaining_steps channel and aborts with a notice message when fewer than two steps remain.
• Reasoning drift / error propagation: Once the model has committed to a false Thought, it interprets subsequent observations so that they fit it. Countermeasure: shorter trajectories, intermediate validation and, for verifiable tasks, an external reflection/check signal (a bridge to the Reflexion pattern).
• Hallucinated tool arguments and hallucinated observations: With ambiguous tool descriptions the model invents arguments; weaker models sometimes also "fabricate" observations instead of waiting for the real tool result. According to research, malformed JSON tool arguments are the most common production error on weaker models. Countermeasure: use structured output or function-calling mode whenever the provider supports it, and phrase tool descriptions precisely.

Added to this is the context bloat already mentioned: every step resends the complete history.

A concrete example: a research agent for a marketing agency

Suppose an agency builds an agent that researches the current headcount of a client company. The loop might run as follows:

1. Thought: "I do not know the current figure and need to look it up." → Action: Search["headcount Company X 2026"] → Observation: hit with a press release.
2. Thought: "The release states a range, I need the exact figure." → Action: Open[URL of the press release] → Observation: full text with "around 1,200 employees".
3. Thought: "That answers the question." → Action: Finish["approx. 1,200 employees, as of 2026"].

Three steps, three LLM calls, a fully logged history. If you set MAX_STEPS = 15, the agent is protected against infinite loops. If it finds nothing, it aborts cleanly instead of burning tokens. It is precisely this traceable sequence that is a real advantage for marketing teams without a developer background: in n8n, for example, the execution view shows every LLM and tool call as an auditable log.

Frameworks and practical recommendations (as of 2026)

ReAct is natively available in all common stacks: LangGraph (create_react_agent), CrewAI (every agent is internally a ReAct agent), AutoGen or the Microsoft Agent Framework (AutoGen is in maintenance mode as of 2026; Microsoft directs new projects to the Agent Framework) and n8n. There, the AI Agent node has run the reasoning-action loop as a Tools Agent via native function calling since n8n 1.82.0; selecting a separate ReAct agent type is deprecated. In LangChain, create_react_agent is moving as of 2026 from langgraph.prebuilt to langchain.agents.create_agent with middleware decorators.

The most important practical insight comes from Anthropic ("Building Effective Agents", December 2024) and Cognition: start with the simplest pattern, usually ReAct, and only escalate when measured failure modes force you to. Modern frontier models handle the reasoning-action loop natively via function calling; explicit ReAct prompting is often unnecessary. What matters then is memory, iteration limits and tracing. Observability tools (LangSmith, Arize Phoenix, Langfuse) are standard practice in production. The EU AI Act requires automatic logging for high-risk AI systems (Art. 12 AI Act); typical marketing and research agents usually do not fall into that category. A persisted trace still pays off for debugging and traceability. If it contains personal data, the GDPR principles of data minimization and storage limitation apply (Art. 5 GDPR): strip PII and set a deletion period.

For agencies and B2B decision-makers

ReAct is the pragmatic entry point for almost any agent project: customer chatbots with CRM and knowledge-base integration, support triage or research tasks. It is quick to set up and fully auditable, and both count in the DACH B2B environment. Cost and latency grow with every step, though. So set hard iteration limits from the outset, use function calling instead of free-text arguments and persist the trace for audit and debugging. If, as an agency, you want to design an agent or have an existing workflow checked for stability, get in touch. We support selection, architecture and compliant implementation.