An action is a step that affects the world or fetches new information: calling a tool, querying a database, sending a message, executing code. A thought is internal reasoning the model writes for itself before deciding what action to take. In the ReAct pattern, thoughts and actions alternate explicitly so the runtime can intercept and execute each action separately.
**An action** in an agent loop is a step that actually does something in the world: calling a tool, querying a database, running code, sending an email, clicking a button. **A thought** is the model's internal reasoning, the text it writes to plan or explain its next move. Both come out of the same LLM call, but only actions cause real effects.
The distinction comes from the ReAct pattern, which made the difference explicit. In ReAct, the model alternates between two kinds of output. First it writes a thought, like "I need to find out the user's recent orders before answering." Then it emits an action, like a tool call to the order database. The runtime intercepts the action, executes it, and feeds the result back as an observation. The cycle repeats: thought, action, observation, thought, action, observation.
Why the split matters. The model's reasoning text is useful for debugging and for guiding its own next decision, but it does not affect the outside world. The action is the part that does. By making the structure explicit, the runtime knows which parts to actually execute and which parts to just log.
In modern function-calling APIs, the action is the structured tool-use block in the response. Everything else (reasoning, partial answers, planning) counts as a thought.
**An action** is a step that affects the world or fetches new information. Calling a tool, querying a database, running code, sending a message, executing a shell command. **A thought** is internal reasoning the model writes for itself before deciding what action to take.
The distinction comes from the ReAct (Reason + Act) pattern, introduced in 2022. Earlier agent designs interleaved reasoning and tool use implicitly; ReAct made the structure explicit. The model alternates: thought, action, observation, thought, action, observation. The thought is plain text. The action is a structured tool call the runtime can parse and execute. The observation is whatever the tool returned. This explicit interleaving made trajectories debuggable and gave the runtime a clean place to intercept and execute actions.
In modern function-calling APIs, the structural distinction is encoded directly. Anthropic returns content blocks of type `text` (thoughts) and `tool_use` (actions). OpenAI returns assistant messages with optional `tool_calls`. Both providers treat the tool-call block as the action; the surrounding prose is reasoning the runtime forwards to the conversation but does not execute. The model itself decides what to emit on each turn: an action means "runtime, please do this"; no action means "I am done" or "I am still thinking."
Why the split matters practically. Actions have side effects, latency, and cost; thoughts are free in comparison. Actions need permission checks, idempotency keys, and audit logging; thoughts only need to be logged for trace inspection. Conflating them is the root of most beginner mistakes in agent design.
Quick reference
TL;DR
An action is a step that touches the world or fetches new information; a thought is internal reasoning. In ReAct, the model alternates them so the runtime can execute actions and let thoughts pass through.
Imagine writing a recipe out loud. 'I think the dough needs more flour' is a thought. 'I add half a cup of flour' is an action. The first one is talking about what to do; the second one is actually doing it. An LLM agent works the same way. When the model writes 'I should search the database for recent orders,' that is a thought: it has not searched anything yet. When the model emits a structured tool call to the database, that is an action: the runtime sees it and actually runs the query. Thoughts help the model plan and explain itself. Actions are what change something or bring back fresh information.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Single-turn completions where there is no tool use and no loop. The thought-action distinction collapses: the model just produces text; there is nothing for a runtime to execute as an action.
Latency-sensitive tasks where verbose chain-of-thought adds unacceptable response time. Thoughts are tokens and tokens are latency; suppress reasoning text or use a non-reasoning model when the budget is tight.
Privacy-sensitive contexts where the model's reasoning might leak intermediate inferences about the user. Thought traces can contain implicit deductions you do not want logged; consider hiding reasoning text from the trace surface or using a model that supports redacted reasoning.
Quick scan · 2–4 min
Everything important, quickly.
5 min: define action as a side-effect causing or information-fetching step, define thought as internal reasoning, walk through ReAct's explicit alternation, show how function-calling APIs encode the split structurally, and explain why the action boundary is the safety boundary.
Real-world examples
Real products, models, and research that use this idea.
1Anthropic's API distinguishes content blocks structurally: type=text blocks are thoughts, type=tool_use blocks are actions. The runtime only executes tool_use blocks.
2the 2022 ReAct paper is the landmark pattern that made the alternation explicit; every modern function-calling API encodes its split structurally.
3Cursor's coding agent prints visible 'thinking' text and then a separate edit_file or run_shell_command action; the editor only executes the action blocks.
4GPT-5.5 with reasoning mode and Claude Opus 4.7 explicitly support extended reasoning text as a separate channel from tool calls, so thoughts can be long without polluting the action surface.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QWhy did ReAct interleaving improve agent performance over a model that emits tool calls without explicit reasoning?
Sign in or go Pro to see the approach for this follow-up.
QHow do extended-reasoning models like Claude Opus 4.7's extended thinking and GPT-5.5 reasoning mode change the thought-action picture?
Sign in or go Pro to see the approach for this follow-up.
QWhy is the action boundary, not the thought boundary, where safety controls belong?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Treating the model's reasoning text as an action. Reasoning is a thought; nothing has happened until the model emits a tool call the runtime executes.