Concept guide · How AI works
An agent is a model using tools in a loop.
The word "agent" sounds like a cleverer chatbot. It is simpler than that and more useful to picture. An agent plans, calls a tool, looks at what came back, decides the next step, and repeats, until something tells it to stop. The loop is the whole idea.
What it actually is
Anthropic's engineering guidance says it plainly: "Agents are typically just LLMs using tools based on environmental feedback in a loop." An LLM, or large language model, is just the model behind a chat assistant. The feedback part is the point. At each turn the model does not guess how it is doing, it checks. The same guidance stresses that during a run "it's crucial for the agents to gain 'ground truth' from the environment at each step (such as tool call results or code execution) to assess its progress."
So an agent is a model that acts, reads a real result, and uses that result to choose what to do next. Take the loop away and you have a single answer. Add it and you have something that can work through a task in steps.
A normal example
You ask an assistant to find the three cheapest same-day couriers for a parcel and put them in a table. A one-shot answer would guess three names from memory. An agent runs the loop: search for couriers, read the results, check which cover same-day, look up a price for each, notice one has no same-day option, drop it and find another, then build the table from what it actually found.
The proof: read the loop step by step
The inspectable object here is the trace: the list of steps the agent took. You can read it like a receipt, and each line is either a real action or a decision based on a real result.
- Plan. Break the goal into a first step: find candidate couriers.
- Act. Call the search tool. This is a real tool call, not a sentence.
- Read the result. Take in what actually came back, the ground truth for this step.
- Decide. One is not same-day, so choose to replace it rather than press on.
- Loop. Repeat act, read, decide until the goal is met or a stop condition fires.
The step that separates an agent from chat is the fourth one. The next move depends on what the last result actually was, not on what the model hoped it would be.
Try it now
- Give an agentic tool a small multi-step job that needs looking things up, such as "check these three facts and list any that are wrong."
- Open the run and read the steps it took, the tools it called and the results it got back.
- Pick one step and check its result against the real source yourself.
The decision this gives you
A loop is worth it when the task genuinely has steps that depend on each other and on real results: look something up, act on what you find, adjust. It is overkill when one good answer would do. Anthropic's guidance is direct about this: "We recommend finding the simplest solution possible, and only increasing complexity when needed. This might mean not building agentic systems at all." If a single prompt answers the question, do not wrap it in a loop.
Treating an agent as more trustworthy because it "did the work itself." A loop repeats actions, including mistakes. A wrong early step becomes the ground truth the next step builds on, so one bad reading can steer the whole run. More steps is more to check, not less.
The human bit
You decide where the loop must stop and wait for you. The guidance builds this in: "Agents can then pause for human feedback at checkpoints or when encountering blockers." Put a checkpoint before anything that sends, pays, deletes or is hard to undo. Read the trace, not just the final answer, because the final answer cannot show you the step where it went wrong.
Where this stops being exact
How much an agent plans ahead, and how freely it acts, varies by tool and setup. A loop needs a way to end: "The task often terminates upon completion, but it's also common to include stopping conditions (such as a maximum number of iterations) to maintain control." Which actions run on their own and which wait for you is a design choice, not a default, and the same loop can be tightly supervised or nearly hands-off depending on how it is wired.