The First Job to Give an AI Agent — and the One to Keep
Ten days ago I wrote about what AI agents actually do. That post was mostly explaining a category. This one is different, because in the space of a week the category turned into things you can open on your phone.
On 7 July, Anthropic brought Claude Cowork to web and mobile. Two days later, OpenAI launched ChatGPT Work, running on its new OpenAI’s latest model family, released alongside ChatGPT Work. It comes in three versions with names you’ll see in menus: Sol, the most capable and slowest; Luna, tuned for speed; and Terra, the balanced everyday option. You mostly don’t need to care — but if a long job is producing sloppy work, moving it to the more capable model is the first thing to try, and if a quick job feels sluggish, the reverse. models. Both are pitched at the same thing: you describe an outcome, and the tool plans the steps, uses your apps and files, and comes back with finished work.
If your company pays for either, this is probably already switched on for someone in your building. So the practical question is no longer “what is an agent.” It’s what do I hand it on Monday — and what do I keep.
What actually changed
Two things, and neither is “the model got smarter.”
It runs when you’re not watching. Cowork works in the background in the cloud even with no device online. You can start something at your desk, close the laptop, get a status update on your phone, and collect the output later. When it hits a step needing human judgment, it sends the question to your phone and waits for you to approve before continuing.
It reaches into your actual tools. ChatGPT Work connects through Plug-ins that let an AI tool read and write in the software you already use — Slack, Teams, Google Drive, SharePoint, email, calendars, CRM systems, project trackers. This is the difference between a tool you paste into and a tool that fetches its own inputs. It’s also why permissions suddenly matter: a connector inherits whatever access you granted it. to Slack, Teams, Drive, SharePoint, email, calendars, CRMs and project trackers. You describe a goal — prepare the quarterly report, research these competitors, organise this folder, build this spreadsheet — and it breaks that into steps, gathers what it needs, and revises as it goes.
Worth knowing about availability, because it’s uneven: Cowork on web and mobile went to Max subscribers first, with desktop still the fullest version (it’s the one that can reach local files and your browser). ChatGPT Work opened to Pro, Enterprise and Edu, with Plus and Business on a phased rollout. Check which tier you’re on before you plan a workflow around it.
One detail from Anthropic’s own usage data is the tell for this whole shift: when they looked at what people did with Cowork, more than 90% of it wasn’t software development. These are no longer coding tools that office workers borrow.
The first job to give one
The instinct is to pick something impressive. Don’t. Pick something boring, bounded, and checkable, because your first run is really a test of your own supervision, not of the tool.
Four criteria:
- It takes you 30–90 minutes and it’s mostly assembly. Gathering, reformatting, cross-referencing. Not deciding.
- You already know what “right” looks like. You’ve done it before, so you’ll spot a wrong answer in seconds rather than believing it.
- Nothing leaves the building unread. Internal, and you’re the reviewer.
- The inputs are things you’re allowed to hand over. If your employer hasn’t told you in writing what may go into these tools, that question comes first — see Write a One-Page AI Policy Your Team Will Actually Follow.
A good concrete first job: “Read the last four weekly updates in this folder, then draft this week’s from the notes in this document. Match the structure of the previous ones exactly. Where a number is missing, leave a marked blank — do not estimate.”
That last sentence matters more than the rest of the prompt. An agent that runs unsupervised will fill gaps to finish the job, and a plausible invented figure is far more expensive than an obvious hole.
The first job not to give one
Anything where you couldn’t tell a good result from a bad one at a glance.
That rules out more than it sounds like. Don’t start with something built on data you can’t verify, or a task whose output goes straight to a customer, or a multi-week project you’d have to unpick if step three went wrong. And don’t start with the thing you personally find hardest — the temptation is real, and the hard cases are exactly where a confident wrong answer slips through, because you don’t have a strong prior for what the answer should be.
Also: nothing where the point is that a person spent the time. Feedback, condolences, an apology, praise meant to land. An agent will produce those, competently, and it will cost you something you can’t get back.
Supervise the seams, not the middle
The pitch is that you don’t have to watch it work. Fine — but there are two moments you do have to be present for, and they’re both at the edges.
The permission prompt. When it stops to ask, that’s the design working. Read what it’s actually asking before you tap approve. A half-glanced approval on a phone is how an agent ends up emailing a customer or overwriting a file you needed. Slow down at the exact moment the tool is inviting you to go fast.
The output. Check facts, not tone — the tone will be fine, and the tone is not the risk. Numbers, names, dates, quotes, anything cited. Run it against one typical input and one awkward one (missing column, unusual case), because the awkward input is where the failure actually lives, and the failure to design against isn’t output that looks wrong. It’s output that looks right.
Then keep the receipt. If a workflow becomes something you rely on weekly, write down what it does, where the prompt lives, and what breaks if it stops working. A saved prompt in one person’s chat history that the team now depends on is a single point of failure with a friendly interface.
The takeaway
The interesting change this month isn’t capability, it’s location. These tools now run without you in the room and reach into your real files, which moves the skill you need from writing a good prompt to scoping a job and checking a result.
So give one a boring, bounded, checkable task this week — something you could have done yourself in an hour and can verify in five minutes. Read the permission prompts properly. Check the numbers, not the prose.
Get that loop right on something small, and you’ll know whether to trust it with anything larger. Skip it, and you’ll find out the expensive way.
Sources: Claude Cowork’s web and mobile launch, 7 July 2026 — Anthropic, TechCrunch, and VentureBeat for the finding that over 90% of Cowork use isn’t software development. ChatGPT Work and GPT-5.6, 9 July 2026 — Forbes and Computerworld. Product capabilities and availability tiers are as reported at launch and will move — check your own account. The task-selection criteria and the supervision advice are mine, drawn from how these failures actually show up, not from either vendor.
Related: What AI ‘Agents’ Actually Do — and What They Can Do for Your Day is the plain-English explainer this post assumes you’ve read. How to Make an AI Agent Without Writing Any Code is the configuration side — the four decisions to make before you hand one a job at all. Build a Reusable AI Assistant for the Task You Do 20 Times a Week is the smaller-scale version of the same idea — worth doing first if handing over a whole multi-step job feels like a leap.