How are AI agents changing work inside Anthropic?

{"text":[[{"start":null,"text":""}],[{"start":8.08,"text":"Welcome back to The AI Shift, our weekly exploration of AI’s impact on the labour market. It’s now a year since we started this newsletter (time flies!), and over that time we’ve dug into the way AI is playing out in all sorts of settings, from radiology to translation and call centres. This week we’re asking: how is AI changing the quantity and quality of work inside AI companies themselves?"}],[{"start":8.58,"text":"Sarah writes"}],[{"start":34.08,"text":"A fun story did the rounds recently about Anthropic posting a job advert for an executive assistant. The role description included — among other things — managing leaders’ calendars, inboxes, travel logistics and expenses. As Seb Johnson of the Scaling Europe newsletter quipped: “Somebody should really build some sort of tech tool that can handle this stuff autonomously.”"}],[{"start":56.72,"text":"It does raise a good question, though, which is whether AI is actually disrupting work inside AI companies as much as those companies claim it will disrupt work everywhere else. As it happens, Anthropic published a richly detailed report into its own workforce this week, which goes some way towards answering that question."}],[{"start":75.32,"text":"The company drew on its internal data on employment levels, AI use, output metrics, employee surveys, internal communication patterns, employee calendar data and hiring decision records, and combined those metrics with 17 employee interviews."}],[{"start":89.92,"text":"Here are three things that stood out to me."}],[{"start":92.32,"text":"First, it does seem that AI is having some impact on the quantity of jobs at Anthropic. That is obviously quite hard to judge in a company which is also going through explosive growth, but the report notes that it hasn’t hired as many people, per dollar of revenue, as the previous generation of fast-growing tech companies did. At five years old, Anthropic says it takes between two to four times as much revenue per employee as Meta and Google when they were five years old."}],[{"start":119,"text":"More specifically, the company has been hiring relatively fewer junior-level employees in technical departments, particularly since January this year, as the capability of its models to do junior-type work has increased. The share of junior-level employees in technical departments at Anthropic has declined from 21 per cent to 14 per cent since early 2025, while the proportion of junior technical hires has fallen from 27 per cent to 9 per cent."}],[{"start":144.48,"text":"As regular readers of the newsletter will know, we’ve been picking up indications of a decline in junior hiring rates in the macroeconomic data, too. There are various competing explanations for that, of which “AI is displacing the need for people with junior-level skills” is only one. But in this case, Anthropic is explicit about AI being the cause."}],[{"start":165.28,"text":"Second, employees at Anthropic seem to be engaged in constant attempts to automate more and more of their own jobs. One employee quoted in the report described an “eternal pressure at Anthropic to move as fast as possible and use Claude for everything”. Speaking of executive assistants, for example, they are (according to this report anyway) busy building AI agents for lower-level tasks so they can focus on “higher-order” activities. A team which used to write stuff like the greetings on Claude.ai, meanwhile, developed a way to guide agents to write the content. The human writers “rotated into agent prompting” and writing model evaluations instead."}],[{"start":204.08,"text":"That brings me to my third takeaway, which is that people seem to be turning from do-ers into managers. And not just regular managers, but managers of hyperactive AI agents, all of which clamour for limited human attention. One engineer compared them to children: “They’re like, I need my juice box opened and it’s an emergency that my juice box is opened. And that is voiced with the same level of concern as the one who has sliced their finger and is bleeding . . . When you are babysitting, every child brings to you concerns with an equal demand for your attention and the various [Claudes] also do that.” All this activity seems to require increased human-to-human co-ordination, so the amount of time people spend in meetings is up. “A break from managing the agents typically looks like jumping into a meeting,” the report says."}],[{"start":251.08,"text":"If you’re wondering whether this sounds like a future of work you’d actually enjoy — well, I’m with you. Some people love to manage and are good at it; others don’t, and aren’t. Even Anthropic seems to acknowledge that it’s pretty tough to work in this intense and constantly shifting way. The report’s authors say it’s important, for example, that employees have stock ownership so they know they share in the financial upside. “At Anthropic, workers codify their expertise for the machine, see their work routines constantly roiled and disrupted, and have seen a large part of their job become speedreading the output of their AI collaborators,” the report says. “They do not do this for free.”"}],[{"start":290.82,"text":"Of course, Anthropic is nothing like most other companies, from its core activity (making the product it is also adopting) to the blistering pace of its growth and the sort of employees it attracts. So — while very interesting — I would be wary of reading this as a preview of the future of work in every company."}],[{"start":308.1,"text":"What do you think, John? Even though Anthropic is quite idiosyncratic, do you think we should draw some wider lessons from its experience so far?"}],[{"start":308.6,"text":"John replies"}],[{"start":318.84,"text":"Thanks, Sarah, lots to chew on here. While I certainly agree we shouldn’t assume what has been true for Anthropic will be true for all white-collar companies, I did find the report really interesting. Deep dives like this that blend data with interviews and reflections add a huge amount of value that isn’t contained in incremental updates on the tasks AI can perform to 50 per cent accuracy."}],[{"start":340.56,"text":"A few things in particular gave me pause for thought. One is the question of what a company is doing when it hires a junior worker. Yes, you’re hiring someone to do research, run calculations and make slides — all of which the frontier models can now do quickly, indefatigably and at a high standard (if you disagree about this, spend some time with Opus 5.5 or Astra). But you’re also hopefully hiring your managers and senior leaders of tomorrow — and it’s a lot less clear AI can replace those. So I do still wonder whether these pullbacks in entry-level hiring are the start of a long-term trend, or may prove to have been a temporary blip that reverses over the longer term."}],[{"start":382.74,"text":"There was also lots to ponder on the implications for education. Anthropic’s hiring managers said they increasingly want evidence that job applicants have demonstrated creative use of AI tools. This crystallises something I’ve been thinking about for a while: although schools and universities will surely need to do much more offline assessment of students (to make sure we’re looking at their knowledge rather than AI’s), they will also need AI-permitted assessments to gauge how skilled students are at getting the most out of AI (since this will now be fundamental to many jobs). I can even imagine assessments where the full chat log between student and AI is submitted — a sort of ‘show your working’ for the AI age."}],[{"start":424.68,"text":"I was also struck by the fact that hiring managers are shifting emphasis towards judgment and autonomy and away from domain expertise. I’m a little unsure how one develops good judgment without domain expertise. There’s a decent amount of evidence that “transferable skills” like judgment and autonomy only really exist as a function of domain knowledge. Someone can make good decisions without supervision because they have a very developed understanding of the domain they’re working on — understanding which is often gained through years of work executing tasks as a junior . . ."}],[{"start":425.18,"text":"Over to you . . ."}],[{"start":461.2,"text":"Calling all accountants! We’re planning a deep dive on AI in accountancy soon, so if you have any experiences you want to share with us, please do get in touch: AIshift@ft.com."}],[{"start":461.7,"text":"Recommended reading"}],[{"start":473.28,"text":"Our San Francisco-based colleagues Richard Waters and George Hammond are doing a live Ask the Expert session today from 8am PT/4pm BST on Anthropic’s IPO and the AI boom (Sarah)"}],[{"start":486.08,"text":"Earlier this month I joined the Jimmy’s Jobs of the Future podcast on YouTube for a wide-ranging conversation about all things AI and jobs. I think it gives a good sense of how I’m thinking about all of this (and more) at the moment, so do give it a listen if that interests you (John)"}],[{"start":null,"text":""}],[{"start":505,"text":""}]],"url":"https://audio.ftcn.net.cn/album/a_1792773027_3591.mp3"}