The Agent Loop

An interactive explainer

What is agentic AI?

Agentic AI is a system given a goal that can choose and carry out steps toward it, use tools or information, and adjust its next steps based on what happens.

The term has no single settled definition. The useful question is what the system can decide and do.

  • goal-directed
  • takes actions
  • learns from results mid-task
Diagram: a feedback loop with five stages arranged in a ring — Goal, Choose, Act, Observe, Adapt — and back to the goal. A glowing pulse travels around the ring, showing that results feed into the next decision rather than flowing one way.

Exhibit 01 · The control loop

Five questions an agent keeps answering

Agents are best understood as a loop, not a pipeline. Select a stage to see what happens there. Use the arrow keys to move between stages.

Stage 01 · Goal

Goal

What outcome is the system trying to achieve?

An agent starts from an objective rather than a complete script. “Find reliable evidence and summarize it” leaves room for the system to work out the steps. The clearer the goal and its limits, the easier it is to judge success.

In the research example“Summarize what reliable research says about screen time and teenagers’ sleep, with sources.”

Worked example: a research assistant

Watch the loop run twice. The first search isn’t good enough, so the system changes approach instead of summarizing too early.

  1. GoalReceives the goal: summarize what reliable research says about screen time and teenagers’ sleep, with sources.
  2. ChooseDecides to start with scholarly sources rather than general web pages.
  3. ActCalls a search tool: “adolescent screen time sleep systematic review”.
  4. ObserveReads the results: mostly news articles and one small study. Not enough to support a summary.
  5. Adapt ↺Judges the evidence insufficient and changes approach instead of writing a summary now.
  6. ChooseChooses a new query aimed at review articles and recognized health organizations.
  7. ActSearches again and opens the most relevant papers.
  8. ObserveFinds review-level evidence from more than one independent source, and notes where studies disagree.
  9. Adapt ✓Judges the evidence sufficient, stops searching, and writes a summary that cites sources and flags uncertainty.

Stopping rule: if it had hit its search limit without enough evidence, a well-designed assistant would report what it found and hand the question to a person rather than guess.

Exhibit 02 · Compare

Three systems, three levels of agency

All three might use the same language model. What differs is who decides the next step and whether results feed back into that decision. Run them a few times: the workflow always takes the same track; the agent’s path changes with what it observes.

Each run is short. The agent’s path is randomized to simulate different results.
Level 1 · Responds

Chatbot

Receives a question and replies. One pass, no actions in the world, no next step.

Ready.

Actions it can take
Chooses its own next step
Uses results as feedback

Next step decided by: the person asking.

Level 2 · Follows steps

Fixed workflow

Follows a predefined sequence. It may call tools and even an LLM, but a developer fixed the order in advance.

Ready.

Actions it can take
Chooses its own next step
Uses results as feedback

Next step decided by: the developer’s code, written in advance.

Level 3 · Decides & adapts

Agent

Chooses some of its own steps, observes what happened, and uses that feedback to decide what to do next, within limits.

Ready.

Actions it can take
Chooses its own next step
Uses results as feedback

Next step decided by: the system at run time, within the tools, limits, and checkpoints it was given.

A single tool call is not enough. A chatbot that always runs one weather lookup before answering is using a tool, but nothing about its path depends on what it observes. Agency is a spectrum: it grows with how many decisions the system makes at run time and how much those decisions respond to results.

Exhibit 03 · Common mix-ups

What agentic AI is not

Labels travel faster than definitions. Each card shows a common claim; reveal the correction to see why it doesn’t hold on its own.

Myth

“It uses an LLM, so it’s an agent.”

Correction

A model that only responds to a prompt may not be an agent.

A language model can power an agent, but on its own it maps input text to output text. Agency comes from the surrounding loop: goals, actions, and feedback.

Myth

“It calls a tool, so it’s an agent.”

Correction

A one-off tool call inside a fixed sequence may be ordinary tool-assisted software.

The question is whether the system decides which tools to use, and when, based on what it has observed so far.

Myth

“It has multiple agents, so it’s agentic.”

Correction

Several agents can still follow a fixed workflow.

A “researcher” handing off to a “writer” in the same order every time is a pipeline with role names. Count the run-time decisions, not the boxes.

Myth

“It learns permanently from every action.”

Correction

Many systems adjust within a task but do not update their model from each run.

Adapting usually means using results in the current context. Lasting change requires separate memory features or retraining, which many agents do not have.

Myth

“More autonomy is always better.”

Correction

Fixed workflows are often more predictable for well-defined tasks.

Autonomy adds flexibility but also cost, latency, and room for compounding errors. A good design uses the simplest approach that works.

Exhibit 04 · Try it

Is this agentic?

Pick an answer for each system. Some cases genuinely depend on design details, and the explanation says which details would change the answer.

01A chatbot answers a history question using only its prompt and what it learned in training.
02A script runs every night: pull yesterday’s sales, fill a report template, email it to the team.
03A research tool searches the web, checks whether the results answer the question, and searches again with a new query if they don’t.
04A warehouse robot sees a box, moves its gripper, checks whether the grip is secure, and adjusts if the box slipped.
05A support assistant looks up an order’s status with one API call, then writes a reply to the customer.
06A “multi-agent” app: a Writer agent drafts an article, then an Editor agent revises it, always in that order.

Answered 0 of 6

Exhibit 05 · Take it with you

A checklist for any “AI agent” product

Skip the label. Ask these six questions about the system itself.

  1. What goal does it receive?

    An open-ended objective leaves room for decisions; a fully specified procedure doesn’t.

  2. Which decisions can it make during execution?

    Choosing tools, order, queries, or when to stop? Or only filling in text inside fixed steps?

  3. What actions or tools can it use?

    What can it read, write, send, buy, or move? This sets both its usefulness and its risk.

  4. Does it use action results to choose what happens next?

    The key test. Do observations change the next step, or does it take the same path regardless?

  5. What are its limits, stopping rules, and human checkpoints?

    Step or budget limits, permissions, approval gates, and when it hands work to a person.

  6. Is success verified in the environment, or does the system merely report success?

    Check the world: did the file change, the test pass, the message send? Don’t rely on a “Done!” message.

The label matters less than the control loop: what the system can decide, what it can do, and how it checks the result.

The one-sentence takeaway.

Sources

Sources and further reading

These sources frame “agent” differently, and vendors do not share one strict definition. This page draws on their common ground: goals, actions, and feedback. It does not present any single definition as official.