Is this the ‘take off’ moment for AI agents?

This article is an on-site version of our The AI Shift newsletter. Premium subscribers can sign up here to get the newsletter delivered every Thursday. Standard subscribers can upgrade to Premium here, or explore all FT newsletters

Welcome back to The AI Shift, our weekly dive into the evidence on how artificial intelligence is impacting the labour market. This week, as US software and professional services stocks tumble over fears of rapid disruption from the latest iteration of AI-powered tools, we ask whether the long-promised but long-missing impact of AI on robust measures of productivity in the tech sector is now here.

John writes

For much of last year, it felt like the AI-and-work story was a standoff between insistent anecdotes of massive productivity gains or impending layoffs on the one hand, and stubbornly stable trends in the hard labour market data. Evidence that AI is behind hiring slowdowns has proved equivocal at best, and as we covered at length in October, the productivity claims were generally refuted by directly measured evidence.

But this week I’ve taken another look at some of the indicators we and others were monitoring a few months ago, and with fresh eyes and fresh numbers, the signals in large-scale productivity data appear to be converging with the anecdotes.

Last September a comprehensive and widely-read analysis by veteran software engineer Mike Judge found essentially no sign of any uptick in software productivity as measured by a range of indicators. These included the volume of code uploaded to projects hosted on the giant coding platform GitHub, volumes of new mobile apps released on Apple’s iOS app store, and new website registrations. I’ve revisited the same measures five months on, and all three show clear upward inflection points by the end of 2025, coinciding with the launch of agentic coding tools such as Anthropic’s Claude Code and OpenAI’s Codex.

Some content could not load. Check your internet connection or browser settings.

By the third quarter of 2025 (the latest data available), the volume of coding additions or updates being ‘pushed’ to GitHub in the US was 30 per cent higher than had it continued on its pre-2025 trend. And all that new code powers new products and services. After years of negligible growth, 55 per cent more iOS apps were released last month than in the previous January. The total number of new website registrations was up 34 per cent year-on-year globally after several years of stability.

Some will doubtless counter that a lot of this AI-produced code is just a new form of ‘slop’. Some of it may be, but the strongest rebuttal to this line of thinking comes in the fact that Claude Cowork itself — Anthropic’s agentic AI platform for non-coders, whose new suite of tools has sparked panic across the digital services sector — was coded entirely by Anthropic’s agentic coding tool Claude Code. So in a matter of months we’ve moved from ‘no solid evidence that AI is boosting productivity’ to ‘AI is now writing fully functioning commercial software that threatens entire industries’ — and the moment when it all changed appears to have been the leap from chatbots to agents.

Sarah, the data looks convincing to me, but am I over-interpreting it? And will what we‘re seeing with code spread elsewhere?

Sarah writes

I think there are two distinct questions here, which are getting somewhat confused in the wider discourse. The first is: are we now seeing a substantial uptick in productivity in the software development realm thanks to agentic AI tools such as Claude Code? The answer to that (thanks to your data analysis) seems to be yes. It’s hard to interpret a 55 per cent year-on-year increase in apps and a 34 per cent increase in website registrations any other way (although, as you note, quantity metrics don’t tell us anything about quality).

The second question is: will we now see a similar surge in productivity across a range of other professions? Consider this recent post on X by Anthropic’s Boris Cherny, who created Claude Code:

“Pretty much 100% of our code is written by Claude Code + Opus 4.5. For me personally it has been 100% for two+ months now, I don’t even make small edits by hand. I shipped 22 PRs yesterday and 27 the day before, each one 100% written by Claude. Some were written from a CLI, some from the iOS app; others on the team code largely with the Claude Code app Slack or with the Desktop app. I think most of the industry will see similar stats in the coming months — it will take more time for some vs others. We will then start seeing similar stats for non-coding computer work also.” [emphasis mine]

For tech people like Cherny, who are experiencing a complete transformation of their profession, it’s probably easy to assume the same is coming for every other type of work. Anthropic’s new tool, Claude Cowork, is explicitly aimed at doing just that. Its tagline is: “Claude Code transformed how developers build software. Now, Cowork brings that same execution power to everyone.”

As we discussed in a previous newsletter about our own experiments with these tools, it’s definitely true that agentic coding capabilities can be useful in professions outside of software development (such as ours!). They might help white-collar workers to do certain tasks faster, from organising files to analysing data, or to embark on projects they simply wouldn’t have been able to do before. (For more on Claude Cowork, Simon Willison wrote a useful review here).

But I think it’s still reasonable to assume that coders’ jobs probably represent the high-ceiling of how transformational these new tools can be in terms of a person’s day-to-day work. And thus, it’s worth taking a breath before extrapolating those software development productivity metrics to the economy at large (and that’s before you consider the likelihood that adoption will also be slower and more cautious in companies outside of tech for security and other reasons).

That said, I don’t think it’s hyperbolic to say that we’re now entering a new phase of AI’s impact on the world of work. If 2025 was the year that everyone banged on about “agentic AI” but nothing really happened (outside of coding), 2026 is the year it’s really going to enter the mainstream.

John replies . . .

Thanks Sarah. I certainly think it’s important to distinguish between the implications for coding vs non-coding white collar work, but I’m not sure the boundaries here are as clear as some might imagine. Ultimately, almost every action that takes place on a computer — searching for information on the web, writing a report, making a slide presentation, sending an email — can be performed by writing and executing code.

As ever, this is not to say that anyone who is paid to do those things is in imminent danger of disemployment. White-collar jobs are much larger and messier bundles of tasks, and even though Claude Code may have written all the code for Claude Cowork, it still needed to be told what to do by human engineers. Even in the most digital-centric occupations, creative thinkers who exercise autonomy in their jobs today stand a good chance of being employed to do the same in the future.

But as swarms of AI agents are let loose to take on more and more tasks over the coming months, I suspect the upticks we have seen in code production will start showing up in a broader range of digital outputs.

Recommended reading

1. Tech analyst Ben Thompson is not convinced that it’s game over for software (John)

2. A fascinating dispatch from our Asia tech correspondent Zijing Wu about China’s “genius classes” and the AI race (Sarah)

Recommended newsletters for you

The Lex Newsletter — Lex, our investment column, breaks down the week’s key themes, with analysis by award-winning writers. Sign up here

Working It — Everything you need to get ahead at work, in your inbox every Wednesday. Sign up here