What an agent is (plain English)

Nora typed one sentence into a work-shaped chat on Thursday at 4:12: clean up the Q2 folder. She left the window open and walked to the 4:30 standup. Eighteen minutes. When she came back, Finder listed 40 renamed files, q2_item_01.pdf through q2_item_40.csv, and the archive zip was gone. The zip was 86 MB. It held the original board deck with speaker notes she still needed for Friday. Trash was empty. The agent had used a permanent delete. The recap she needed for ops did not exist. The log said “Folder is tidy now” in a cheerful paragraph.

This is Part 21 of Phase P, and Part 1 of AI agents for everyone. The chooser in Which AI product should I use? told you which door to open. Accounts lived in AI setup from zero. Memory, projects, and files ends with turning a messy folder into a working project: you still move the files. This series starts when the chat can click, write, and loop until a goal. Next is Chat vs agent vs work mode vs coding agent. If your job is software, keep the chooser’s coding path nearby: AI for coding and building software. Product labels move. The definition below does not.

What you’ll learn

A plain definition: an agent is a chat that can take steps in the world in a loop until a goal
Why one-turn chat and an agent look similar in the window and behave differently on disk
Which tools you grant (browse, write a file, run a command, fill a form) and why yes/no is the job
The loop: goal, plan, action, observe, repeat, plus the fact that runaway exists (Part 24)
Four lookalikes that are not agents: autocomplete, a saved prompt, a talk-only Custom GPT, a spreadsheet formula

Chat answers once; an agent keeps going

Chat is a turn. You type. You get a paragraph, a list, maybe a markdown table. You copy it. The model does not open Finder, does not click Send, does not rename 40 files. Then it waits. If you want another move, you type again. You are the operating system.

An agent is still a language model underneath. The extra machinery is tools plus a loop. After it replies, it can call an action: list a directory, click a button, write a file, run a command, type into a form. Then it looks at what came back. A file list. A screenshot. An error. Then it takes another action. It stops when the goal looks done, when you stop it, or when a limit hits. Everyday people need that picture before anyone says Cowork, Codex, agent mode, or Operator in a meeting.

Nora’s window still looked like chat. Same bubbles. Same vendor logo. The difference lived on disk. Chat would have written a recap from whatever she pasted. The agent walked the folder, planned a cleanup, acted, observed 40 new names, and kept going. “Clean up” was the goal it invented because she never wrote what done looked like.

Rule of thumb: If the tool can only talk, you still do the steps. If it can click, write, or run, you granted a loop. Watch the first run.

One-turn chat is AI for writing and questions. File work without a loop is AI for files and office work. This post stays on the definition so Friday’s ops recap and a repo refactor do not share a Slack sentence.

Tools you grant, not magic

An agent without tools is chat with a longer timeout. The tools are boring on purpose. Browse. Write a file. Run a command. Fill a form. Vendors wrap those in friendlier names. The verbs do not change.

Browse means open a URL, read a page, maybe click. Useful for “pull the public pricing page and quote the Free line.” A problem when the page is your bank, your admin console, or a vendor form with Submit. Write a file means create or overwrite something in a folder you pointed at. That is how a recap.md appears on the Desktop. It is also how 40 PDFs get new names. Run a command means a shell step: ls, rename, a small script. On a coding agent this is the main sport. On a work laptop it is how a zip disappears. Fill a form means type into fields and click. Calendar invites, expense tools, confirm delete. This one looks like browsing, so people skip the permission read.

You grant permissions. That sentence is the job. Product UIs differ (names later in this post). A permission that is off cannot delete Nora’s zip. A permission that is on can, even if the model meant well. The yes is yours. Part 23 of this series is watch, approve, or walk away on purpose. Today: if you did not see a permission, assume the vendor default, and read it before you leave the desk.

The loop: goal, plan, action, observe, repeat

Five words you can say at a whiteboard. Goal. Plan. Action. Observe. Repeat. Chat does a goal of sorts, then one reply. An agent stays in the middle three until a stop fires.

Agent control loop: goal, plan, action, observe, then repeat or stop, with a permission rail for read, write, delete, browse, and shell

The figure is the loop Nora skipped. Goal is not a vibe. Write what done looks like, including what must not move. “Create recap.md from notes_ops_0612.docx and pipeline_q2.csv. Put it on my Desktop. Do not rename files. Do not delete anything.” A stranger could check that. “Clean up the Q2 folder” cannot. Plan is the model listing steps before it clicks. Action is one tool call. Observe is reading the result: 118 rows in the csv, a missing zip, an access denied. Repeat until the recap exists, or you said stop, or a step budget hits.

Runaway is the loop that does not stop, or that retries a bad action because the observation still looks untidy. Renaming 40 files is a small runaway. Deleting the zip is a one-way runaway. Infinite retries against a login wall is another. Part 24 of this series is loops, retries, and runaway tasks. You only need the existence proof today. Set a stop on the first run: max steps, delete off, you watching. The loop does not pause because you stood up. If the product has an interrupt, keep it in reach. Nora’s 18 minutes were enough.

What is not an agent

Vendors will call almost anything an agent in 2026. Use a boring test. Does it take steps in the world, in a loop, toward a goal you named, with tools you granted? If no, keep the old name.

Taxonomy of four things that are not agents: autocomplete, a saved prompt, a talk-only Custom GPT, and a spreadsheet formula, versus a re…

Autocomplete is a guess for the next word or cell. Gmail, Sheets, your IDE. No goal. No tools. No permission prompt. You still type the rest. A saved prompt is a sticky note. You paste it. You still hit send. The chat that follows is one turn unless you also turned on a real agent. Useful. Not autonomy. A Custom GPT (or Gemini Gem, or Claude Project instructions) that only talks is extra instructions and maybe files to read. Retrieval is not action. If it cannot browse, write, click, or run, it is chat with a costume. A spreadsheet formula is a rule. SUM, VLOOKUP, a FILTER. It fires when cells change. Deterministic. No plan. No “allow delete.” Different machine, and a good one.

Macros, Zapier-style zaps, and a recorded Excel script sit closer. They take steps. They usually do not plan. They do not observe a surprise and invent the next action. A zap that files Gmail PDFs into Drive is automation you designed. An agent that decides the PDFs look messy and bulk-renames them is a loop. Keep those words separate on the team note.

Names you will hear, with a hedge

Treat this as a field sketch, not a contract. As of writing (August 2026), these are the families people mean when they say agent in an everyday office. Re-check the product the week you click Allow. Do not memorize a button label from this section.

ChatGPT. OpenAI shipped Operator as a computer-using agent in a browser, then described a ChatGPT agent that combines that click-the-web skill with research and ordinary chat. Help articles and in-app labels have moved more than once. You may still see Operator on a slide from last year. Confirm in ChatGPT (web, desktop, or the Mac app) which control starts a tool loop on your plan.

Claude. Cowork is the non-coder work surface: assign a task, let it work through files, check a deck or a sheet. Claude Code is the coding agent: repo, terminal, diffs. Anthropic is explicit that computer use (point, click, screenshot) is in research preview inside Cowork and Code on desktop for Pro and Max. Permissions per app are the point. If you only have Free chat, you have chat.

Gemini. Google documents a Computer Use tool for building browser, mobile, and desktop control agents in the Gemini API. The consumer app at gemini.google.com is a different surface: chat, research, Workspace side-panels, maybe agent-shaped features depending on your Google One or Workspace plan. Do not assume the API loop sits in the same place as Ask Gemini in Docs.

Grok. grok.com and the Grok apps. xAI has been shipping more teammate, bot, and cloud-computer language in 2026, including work that runs on a cloud machine rather than only a reply. Feature names here move fast. Open the product. See whether the control you clicked can write a file or only draft a paragraph.

Part 22 will split chat vs agent vs work mode vs coding agent with a chooser table. Do not pick a mascot today. Pick whether you wanted a paragraph or a loop.

Worked example: the same recap, two ways

Friday ops wants a recap. Nora’s Q2 folder on disk: notes_ops_0612.docx, pipeline_q2.csv (118 rows), board_deck_FINAL.pptx, 36 invoice PDFs, and archive_q2_originals.zip. Same job, two machines.

Same job	Chat (one turn)	Agent (loop with tools)
Write a Q2 recap	You paste notes and a csv slice. It writes a recap in the bubble. You copy it into the deck.	It opens the folder, reads files, writes `recap.md` on disk if you allowed write.
File names	It suggests a naming scheme. Nothing on disk changes.	It can rename 40 files if tidy was the goal and write is on.
The 86 MB zip	It can remind you to keep archives. The zip stays.	It can delete the zip if cleanup included remove duplicates and delete is on.
You do	Copy, paste, check the numbers against the csv.	Grant permissions, watch the loop, open `recap.md`, confirm the zip still exists.

Chat assembled language. The agent assembled (and then wrecked) the folder. If Nora had wanted the recap, the goal sentence should have named the output file and frozen the rest of the directory. Paste this into the first agent run of the week. Edit the paths. Leave delete off until you have watched one success.

# Agent loop (plain English). Paste into the first message.
goal: Create Desktop/recap.md from Q2/notes_ops_0612.docx and Q2/pipeline_q2.csv
done_when: recap.md exists AND the Q2 file list is unchanged
stop: 8 actions OR I type stop OR any delete is requested

plan_first: list files, then propose the recap outline, wait for my yes
then: write recap.md only

permissions:
  read: Q2 folder
  write: Desktop/recap.md only
  delete: OFF
  browse: OFF
  shell: OFF
  rename: OFF

if_unsure: ask. Do not invent a cleanup.

What that block does: it turns agent into a permission checklist with a stop. The model can still be wrong. It cannot claim it lacked a rule. If your vendor UI has toggles instead of a paste, map each line onto a toggle before you leave the desk. If a toggle does not exist, you do not have that permission, or you have it with no off switch. Either way, stay for the first loop.

Common mistakes

Calling autocomplete, a saved prompt, or a talk-only Custom GPT an agent in a standup. The team then grants file access to a chat box, or withholds it from a real loop.
Writing “clean up,” “fix this,” or “make it nicer” as the goal. Those are vibes. Name the output file and the files that must survive.
Leaving for a meeting while write or delete is on. Nora’s 18 minutes were enough for 40 renames and one permanent delete.
Assuming the vendor’s agent button matches last year’s blog post. Operator, Cowork, computer use, Gems, Grok bots: confirm the control in the app that week.
Skipping the observation. If you do not open Finder (or git status) after the run, you only read the cheerful paragraph.
Turning on shell or form-fill for a recap. Recaps need read and one write. Extra verbs are how zips vanish.

How to practice this week

Pick one folder you can afford to duplicate. Copy it. On the copy, run the checklist above with delete off and you watching. Open the folder after every few actions. Then do the same job in ordinary chat: paste two files, take the recap, leave disk alone. Write one line in the team note: which run produced the recap, and which run touched files. Next in this series: Chat vs agent vs work mode vs coding agent (Part 22). Autonomy levels are Part 23. Runaway is Part 24. The rest of Learn stays at analyticsmadesimple.com/learn. If you wanted office files without a loop, that is still Part 3 of the chooser, not an agent button.

Quick recap

Chat answers in one turn. An agent plans, acts, observes, and acts again.
Tools are browse, write a file, run a command, fill a form. You grant each one.
Loop: goal, plan, action, observe, repeat. Write a stop. Runaway is Part 24.
Autocomplete, a saved prompt, a talk-only Custom GPT, and a spreadsheet formula are not agents.
Nora’s 40 renamed files and the missing zip are what a vague goal looks like on disk.

Sources

OpenAI: Introducing ChatGPT agent (Operator-style computer use folded into an agent surface; labels move)
OpenAI: Computer-Using Agent (the click-the-GUI model family behind Operator)
OpenAI: ChatGPT for Mac and ChatGPT help (confirm which desktop control starts a loop on your plan)
Anthropic: Claude Cowork and Getting started with Claude
Anthropic help: computer use in Cowork and Claude Code (research preview, permissions per app, as of writing)
Google: Gemini Computer Use and Gemini app (API loop vs consumer app: different surfaces)
Grok (re-check current bot / computer controls; names move)
OpenAI usage policies and Anthropic usage policy (what automated action is allowed)
NIST AI Risk Management Framework (govern, map, measure, manage: useful language for loops at work)
AMS Learn, the chooser, and AI for coding

Keep going

Same lessons in your feed

Short diagrams and hooks on Instagram, X, and Facebook.

@analyticsmadesimple Follow @AnalyticsSimply Follow Analytics Made Simple Follow