No Meeting, No Manual.

I have been watching people use the new agents. Claude Cowork. ChatGPT Work, which is Codex grown up and moved into the desktop app. Gemini Spark. Grok Bot. Different companies, different models, different price lists, and for the purposes of this piece the differences do not matter.

Ask a person why they like the one they use and you will not hear a benchmark. You will hear the same two answers, in nearly the same words, from a marina owner and a paralegal and a fourteen-year-old. It does it right away. And, I didn’t have to learn anything.

Both answers deserve more attention than they are getting, because together they explain why this wave of software is being adopted faster than anything that came before it, and they point at what the companies building it should protect.

01 · IMMEDIATE GRATIFICATIONIt does it right away.

Think about what happens when you ask a human coworker for something. Not a hard thing. A sheet, a summary, a first draft. You ask. They ask a clarifying question. Someone suggests a call. The call becomes a plan. The plan gets a document. The document gets comments. Two people disagree about scope. A week later you have a meeting about the thing, and the thing itself still does not exist.

None of that is laziness. It is coordination, and coordination is how organizations keep one person’s bad idea from becoming everyone’s problem. But from the seat of the person who asked, it feels like waiting. It is waiting.

Now ask the agent. You type the sentence. It starts. A few minutes later there is a sheet, or a draft, or a folder full of renamed files. No call. No plan. No comments. No week.

Figure 01

The same request, two coworkers

Asking a coworker Tuesday 9:14 → the following Monday

• Can you pull the Q3 numbers into a sheet by region?
• Sure. What format? And do you want it by booking date or ship date?
• Let’s grab 30 min Thursday to align on scope.
• Thursday
• Good call. I’ll write up the plan and share it.
• Friday
• Doc is shared. Priya left comments on the region definitions.
• Following up on this.
• Monday
• Meeting invite sent for tomorrow to close it out.

Asking an agent Tuesday 9:14 → 9:17

• Pull the Q3 numbers into a sheet by region.
• On it. Using booking date, matching last quarter’s sheet.
• 9:17
• Done. Sheet is in the folder, with a summary tab on top.

People love this. Not appreciate. Love. Every consumer technology that has ever won did so by shortening the distance between wanting and having. One-click buying. The song that plays before you finish saying its name. A car that appears at the curb. Work software never offered that, because work software was built for the organization, and the organization runs on deliberation. The agent is the first work tool built for the person who wants the thing.

02 · NO BEHAVIORAL CHANGEI didn’t have to learn anything.

The second thing is stranger, and I think more important.

Every productivity revolution before this one charged a training tax. The personal computer required you to learn files, folders, and a mouse. The spreadsheet required cells and formulas and a new way of thinking about a column. Enterprise software required weeks of classroom time and a consultant living in the building. CRM required a discipline most salespeople never adopted, which is why most CRM data is fiction. Slack required a new etiquette. Each of these asked you to change how you behave before it would give you anything back, and adoption moved exactly as fast as people were willing to change.

The agent asks for nothing. You text it. You text it the way you text a person: a sentence, a little context, maybe a screenshot. Gemini Spark has an email address. Grok Bot sits in a roster next to your chats. Cowork opens as a conversation. There is no new system to learn because the system is the one you already use for every human in your life.

Figure 02

The training tax, by generation of work tool

1. Personal computer 1980s

   Files, folders, a mouse, and the idea that a document lives somewhere.

2. Spreadsheet 1980s–90s

   Cells, references, formulas, and a new way of reasoning about a column.

3. Enterprise software 1990s–2000s

   Weeks of training, a consultant in the building, and a process redesigned around the system.

4. CRM 2000s

   A data-entry discipline most users never adopted.

5. Team chat 2010s

   Channels, threads, and an etiquette for when to use which.

6. AI agent 2026

   A sentence, typed the way you already text a person.

That is ChatGPT’s real legacy, and it was not the model. In November 2022 the literate world learned, more or less in one weekend, that you can type a sentence into a box and get work back. Every agent shipped since has inherited that trained behavior for free. The training tax was paid once, by everyone, years before the products that needed it existed.

03 · TOGETHERWhy the two together matter.

Either of these alone would be a good feature. Together they are the most frictionless adoption in the history of work software. Zero wait, zero training. Wanting to having, with nothing in between.

Notice what is not on the list. Nobody says the model is smart. They assume it. Intelligence is the enabler that makes the other two possible, the way the engine enables the car.

Nobody ever bought a car because they loved the engine.

That has a consequence for the companies building these things. The winners will not be the ones with the best benchmark. They will be the ones who protect the two things people actually love. Every added approval step, every planning mode that wants to discuss before it does, every new panel or workspace or dashboard to learn, spends one of them. Some of that spending is necessary; a bot that moves money should stop and ask. But every product team should know which currency it is spending, and how little of it users are willing to give up.

Already true

All four agents are text-first and act on the request without a planning ritual. The people using them describe the experience in these two terms, unprompted, across ages and occupations.

What has to happen

The loop has to stay short as the tasks get longer, and the interface has to stay a thread you text. The moment an agent needs its own console, or a scoping conversation before it starts, it becomes a coworker again, and the coworker was the problem.

Where I am probably wrong

The meeting was doing something. Discussion, planning, and arguing are how organizations catch a bad ask before it ships. If instant, unquestioning agents let confident wrong work out the door at scale, people may learn to want the friction back, and the products that add a deliberate pause will win.

I would also be wrong if “no new behavior” turns out to be true only for the asking. Giving an agent the access and context it needs, the connectors and permissions and folders, is itself a behavior change. Today it is deferred to whoever sets the thing up. If it lands back on the user, the tax returns.