Describe the typical agent execution loop step by step.
Free · 2 AI evals / day
An **agent loop** is the engine that turns a one shot LLM call into a multi step problem solver. Instead of asking the model for a single answer, you put it inside a loop that runs until the task is finished.
The steps repeat in roughly this order. First, **observe**, the agent reads the current state, which starts as the user's request and grows over time. Second, reason, the model decides what to do next and which tool to use. Third, act, the runtime executes that tool call, for example a search, a database query, or a code snippet. Fourth, observe again, the tool's result is appended to the agent's context. Fifth, check, the model decides whether the goal is reached.
If the goal is reached, the agent emits a final answer and exits. If not, the loop runs again with the new observation in context. A guard rail caps the number of steps and the total cost so a confused agent cannot spin forever.
The key insight is that the model never plans the whole trajectory up front. It plans one step, sees what happened, and then decides the next step from the actual result.
An **agent loop** is a five step cycle that wraps an LLM into an autonomous task runner. The loop is: observe state, reason about the next action, choose and call a tool, capture the observation, then decide whether to stop or continue.
The canonical iteration looks like this:
$$\text{state}_{t+1} = \text{state}_t \;\cup\; \{\text{action}_t, \text{observation}_t\}$$
State is the running transcript of everything the agent has seen so far. It includes the original goal, every prior tool call, and every tool result. At each step the model receives this transcript and emits either a tool call or a final answer.
Termination has three exits: the model emits a final answer, a step budget (typically 10 to 40 iterations) is exhausted, or a cost or wall clock budget is hit. Production loops also include retry logic for transient tool failures and a circuit breaker for repeated identical actions, which is the classic symptom of a stuck agent.
Quick reference
TL;DR
An agent loop is observe, reason, act via a tool, observe the result, then repeat until the model signals done or a step or cost budget runs out.
Picture a person solving a puzzle they have never seen before. They look at what is in front of them, think about the next move, pick up a tool (a calculator, a notebook, a search bar), try something, and then look at what happened. If they are not done, they think again and pick another move. They keep going like that, step by step, until the puzzle is solved or they decide to stop because they have run out of time, energy, or attempts. An agent works the same way. The model is the thinker, the tools are the calculator and search bar, and the loop is the patience to keep going round by round instead of solving it all in one shot.
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.
The execution path is fully known at design time and never branches on intermediate results. A fixed DAG or LCEL chain gives bounded latency and cost without paying for unused dynamic routing.
Sub-100 ms latency budget per user request. Each loop turn typically adds a full LLM round trip, so even three turns blow past tight latency SLAs.
The task can be answered from a single retrieval plus one generation. Plain RAG with no loop is cheaper and more debuggable when the answer fits in one shot.
High stakes domains where every tool call must be audited and approved. Autonomous loops are hard to certify, a human in the loop step function fits better.
Quick scan · 2–4 min
Everything important, quickly.
Walk through the five steps in order, name the state accumulation pattern, contrast ReAct with function calling, list the three termination exits, and close with the repeated action and observability concerns that separate prototypes from production loops.
Real-world examples
Real products, models, and research that use this idea.
1Claude's computer use mode runs an agent loop where the tool is the desktop itself. Each turn observes a screenshot, reasons, and emits a click or keystroke.
2LangGraph's ReAct prebuilt agent wraps OpenAI or Anthropic function calling in a state graph where each node is one loop iteration with full trace logging.
3OpenAI's Swarm framework runs lightweight agent loops with explicit handoff actions, letting one agent yield control to another mid-loop when a subtask needs a specialist.
4Cursor and Cline editors run code modification agents whose loop observes file diffs, reasons about the next edit, calls a file write tool, and reruns tests until the suite passes.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QHow would you detect and break a loop where the agent keeps making the same failing tool call?
Sign in or go Pro to see the approach for this follow-up.
QWhat changes in the loop when you move from a single agent to a multi-agent team with handoffs?
Sign in or go Pro to see the approach for this follow-up.
QHow do you keep the context window from blowing up over a long run?
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
Describing the loop as a one shot plan then execute. Real agents reason between every action so the next step can depend on what the last observation actually returned.