Between April and August, four of the biggest names in AI shipped the same product. Anthropic put Claude Cowork on the desktop, then moved it to the cloud so a session keeps running after you close the laptop. OpenAI shipped ChatGPT Work, an agent you hand a goal and let run for hours until a finished document comes back. Amazon added autonomous agents to Quick, its answer to the same problem. In August, xAI launched Grok Bot, a set of always-on agents that each run on their own cloud computer and sign into the apps you already use.
Four labs, one season, one idea: an agent that does the work inside your tools instead of answering questions in a chat window. When Microsoft, Google, OpenAI, Amazon, and now xAI all ship the same shape in the same year, it stops being a trend and becomes the new shape of the tool. The shape change asks something different of the people using it.
We have had a longer runway on these tools than most, through beta access and the client work that keeps us in them every day. So we will say plainly what we are still working out ourselves: the way to think about this is not fully settled. But it is coming into focus, and the outline is clear enough to be useful now.
Knowledge work has always been two things.
Before anything about AI, put your own week into two columns.
One column is the recurring work. Same shape, every week, month, or quarter. The roll-forward report, the month-end close, the Monday status summary, prospect research before a call, renewal tracking. You know the steps before you start because you have done them before.
The other column is the one-off problem solving. The work in between, never the same twice. Come up with a new pricing model. Work out why a number moved. Build the case for an investment. Plan next year's strategy. Respond to something nobody planned for. You do not know the steps going in, because the shape is new every time.
Almost everything a knowledge worker does lands in one of those two columns. This new generation of tools changes both, in different ways, and the change is easier to use once you can see which column you are in.
What the chat window actually did.
Since late 2022, the chat interface has been how most people reach this technology, and it helped with both columns.
On recurring work, it meant writing a good prompt once and reusing it, saving it in a project or a custom GPT, or returning to the same conversation. If you wanted the work to run on its own, you handed it to someone who could build in n8n, Zapier, or Power Automate. That was a separate tool and a separate skill set, which is why so much recurring work stayed manual.
On one-off work, the whole game was managing context. You uploaded files, connected sources, and tried to get as much of the relevant knowledge into the window as it would hold. You organized the material before you could use it, asked your question, and got one answer back. Then you pasted that answer somewhere and started editing. Nothing durable was left behind. The race was context: more of it, from more places, into a single window.
Both approaches worked. Both are about to look like the slow way.
What changed under the hood.
Three capabilities separate this generation from the chat window, and naming them explains everything that follows.
First, these tools have access to your file system. Point one at a folder and it reads, edits, and creates real files where the work already lives. The limit stopped being what fits in the context window. It is now what you choose to give it access to, which can be years of documents, transcripts, and spreadsheets you never have to upload or trim.
Second, they plan the task before they run it. Instead of cramming everything into one turn, the tool breaks the job into steps, writes the plan down, and works through it while you watch. You can correct the plan before it executes. That is the difference between a long answer and work you can check.
Third, they run on a schedule. A detailed plan written in plain language now does what a flow chart used to do, and you can point it at any hour or event and tell it to repeat. Changing what it does means rewriting a sentence, not rebuilding an integration.
For recurring work, levels one and two are collapsing.
We talk about solutions in three levels. Level one is solving a problem inside the tool you already have, with prompts, projects, and habits, no build required. Level two is wiring the steps together with an automation platform so the work runs on its own. Level three is custom development, purpose-built software for when nothing off the shelf will do.
For years there was a real wall between level one and level two. Level one was something you could do yourself. Level two required a separate platform and someone who knew how to build in it. That wall is coming down. The tool now connects to your systems, plans the steps, schedules the run, and works through the job itself. A lot of what used to require a build in n8n or Zapier can now be described as a detailed prompt and handed to the same tool you were already using. Not all of it, but enough that the person with the problem can often solve it without waiting on someone who builds.
Level three still stands apart. The gap that is closing is between using a tool and automating with one.
This is the moment to look backward before you look forward. If your team has spent a year surfacing ideas and setting some aside as not yet possible, pull those back up. Some were only out of reach because of what the tools could not do yet. Then take an honest inventory of your own recurring work, because you cannot hand off what you have not named. Picture one standing instruction, written once in plain language: review this folder and the last three reports, pull the numbers from Slack and email, do a quick search on these topics, produce a report, a deck, and a dashboard on brand, and post the link to my manager when it is done. Every piece of that exists today. What is new is writing it once and letting it run.
For one-off work, stop stuffing the chat and build from a corpus.
The bigger change is on the other column, and it is a genuine shift in how you solve a problem you have never solved before.
The old way was to push everything into the chat and pull one answer out. The new way is to give the tool a body of source material and let it build the work along the way. It scopes the question, sweeps the sources, distills them into pieces, and assembles the output from those pieces instead of from a single guess. The step that did not exist before is the middle one, where the tool produces real intermediate work product: a source map, a findings memo, a data extract, draft sections. You can open every one of those, correct them, and reuse them.
Your job moves from prompting to conducting. You decide what gets made, in what order, and from what. This works now in part because file output finally holds up. Spreadsheets keep their formulas and tabs. Documents land formatted and ready to send. Decks and dashboards come out as real artifacts you can open. The tool edits the file you already have instead of making a new one. Reliable file output is what turns a good answer into finished work, and it was the flakiest part of these tools until this generation.
The method is a folder, not a chat.
Here is the setup we hand teams, and it is the same every time. A working directory with three folders and one file.
pricing-model-2027/
├── agent.md
├── reference/
├── work/
└── outputs/
The file, agent.md, is the one that makes the rest work. It is not a script of steps to run. It says what the project is, how you want to work together, and the standards to hold: bring me options instead of one answer, ask before making a decision that is mine, cite the source file for every number, use our writing style and brand. The tool reads it automatically at the start of every session in that directory, so the project's memory and standards carry forward and the next session starts where the last one ended.
reference/ is everything the work is made from, dropped in as it arrived and left alone. Proposals, budgets, board decks, transcripts, data exports. No curating, no uploading, no trimming to fit. The tool reads across the whole folder, and the folder keeps growing as new material shows up.
work/ is where the thinking gets written down. The scorecards, models, and summaries the tool builds as it works through the material. This is the folder that did not exist in the chat experience, and it is what makes the output checkable. You read the reasoning here and catch it before anything reaches the deliverable, and later steps build on these files instead of re-reading the whole corpus.
outputs/ holds only the finished pieces. The deck for the room, the one-pager for the pre-read, the model finance will open. Nothing in progress lives here, every claim traces back through work/ to a file in reference/, and everything is on brand because the standards in agent.md apply to all of it.
Four places for things to live, and no ambiguity about which is which. Come back in three months and the reasoning is still there, not buried in a chat history. The directory is a knowledge store now, and it is portable: point any of these tools at it and they pick up where you left off. That is the real change to internalize. You are no longer stuffing context into a conversation and hoping. You are building a durable place where the work accumulates.
The seat is now, not later.
None of this is finished, including our own thinking about it. The tools are days and weeks old, the pricing and access questions are unresolved, and the best patterns for working this way are still forming. We would rather be in the room while it forms than read about it a year from now, and that is the posture we would encourage for anyone whose work lives in those two columns.
The chat window was how we all learned to use this technology. It is not where the work is going. The work is going into your files, onto a schedule, and into a directory you keep. The people who set themselves up for that now will spend the next year getting good at it while everyone else is still pasting into a box.
Sources
- Claude Cowork launch and cloud expansion: Fingerlakes1, Claude Cowork AI Agent Launches
- ChatGPT Work: Bloomberg, OpenAI unveils ChatGPT Work agent to field tasks for hours
- Amazon Quick autonomous agents: AWS, Amazon Quick announces autonomous agents
- Grok Bot: Unite.AI, xAI launches Grok Bot, always-on AI teammates with their own cloud computers