Not a Coding Tool

Anthropic just published data on 1.2 million Claude Cowork sessions. Software development, the work these agents made their name on, was 8.7% of them.

Cowork is the thing Anthropic ships to take a whole task off your plate and hand back a finished file, not a chat window you keep prompting. In Anthropic's own breakdown of what people run through it, the top of the list wasn't code. Business process and operations led at 33.4%, content and copywriting followed at 16.4%, and software development came in third at 8.7%. Anthropic calls the work at the top "the work around the work," the reconciliations and status updates and slide decks that live in no one's job description and eat everyone's week.

The numbers arrived the same week Anthropic pulled Cowork off the desktop. It pushed the agent onto the web and phones on July 7, first for Max subscribers, so a task started at a desk can finish after the laptop's closed. Two days later OpenAI launched ChatGPT Work, an agent that reaches into your connected apps to produce the reports, spreadsheets, and slide decks you'd otherwise assemble by hand. Both labs put the same kind of product on the market within days of each other.

The category started as the AI coding assistant, and the buyer was engineering. The budget sat under a VP of Engineering, the seats went to developers, and success got counted in merged pull requests. Anthropic's own sample says that description now fits under a tenth of the work. Fold in DevOps and infrastructure at 7% and everything an engineer would recognize as theirs still doesn't clear a fifth of sessions. The rest is finance closing the month, an HR team building onboarding checklists, whoever in comms owns the Monday update, and the operations person quietly fixing a spreadsheet no one admits to.

The labs' marketing hasn't caught up to that yet. The same week the 8.7% number went out, Sam Altman was selling GPT-5.6 as 54% more token-efficient than Anthropic's model on agentic coding. The headline benchmark for the engine inside ChatGPT Work is a coding score. That pitch still points at the developer, months after the usage left the IDE.

The mobile move reads like a convenience feature and isn't. A desktop agent works while you watch it. What's new is the run you don't attend, the 6 a.m. task Anthropic uses as its own example, where Cowork reads overnight through your email and meeting transcripts, assembles the client briefing, and has the reply written and waiting by the time you wake up. This reads less like a faster tool than like a coworker who holds your inbox and calendar, works a shift you're not present for, and hands back finished work instead of suggestions.

The fair objection is that this is a rename. A chatbot could already draft a memo, and wrapping one in a task queue and a phone app doesn't make a new category. That holds until you look at what completes without supervision. A chatbot waits for the next prompt. Cowork and ChatGPT Work take a multi-step job, choose the steps themselves, and return a deliverable while you're in a meeting. Whether you have to be in the room is the whole difference between a tool and an employee.

So the decision in front of an operator this quarter isn't which model writes the best code. It's who inside the company is about to be running agents no one watches run. A coding tool had one owner and a contained blast radius. An agent that finance, HR, and comms all aim at their own unglamorous work has neither, and the week that data went public is the week it stopped being engineering's call to make.

What to Do With This

If you're an individual or a small team, take one weekly task that's really just assembly, the status report, the numbers roll-up, the recurring client brief, and hand it to a scheduled agent run instead of building it fresh on Monday. Judge it on the finished artifact, not on the live demo.

If you lead a team that already pays for AI seats, check where those seats sit. They're probably with engineering, because that's who the tool was sold to. The 33% is in finance, operations, and comms, so that's where the next seats belong.

If you decide what the whole company runs, write down who signs off on an agent's output before it leaves the building. These tools now produce finished work unattended, and "the agent drafted it" can't stand in for "a person checked it." Set that rule before the first unreviewed deck reaches a client.

Want More Than This Newsletter?

Alcreon publishes a daily AI briefing, long-form dossiers, and an analysis feed for the teams actually shipping AI in production. This newsletter is one read out of the full library.

Read the daily feed or browse the editorials.

Also on the Radar

SpaceXAI Ships Grok 4.5 at Opus-Class Claims and Well Under Half the Price

SpaceXAI shipped Grok 4.5 on July 8 at $2 per million tokens for input and $6 for output, undercutting the $5 and $25 Anthropic lists for Opus 4.8. Elon Musk called it Opus-class at lower cost, built for coding and agentic work. The frontier tier keeps getting cheaper faster than most procurement cycles can re-run.

The UN Opens a Governance Dialogue as Its Science Panel Warns on Catastrophic Harm

The UN opened its Global Dialogue on AI governance in Geneva on July 6, days after its new scientific panel warned that science can't yet guarantee advanced AI won't cause catastrophic harm. No binding rules came out of it. For operators the signal is direction rather than a deadline, and the compliance surface for frontier models is only going to widen.