Hanging out with friends last Saturday, I asked someone who manages a team in IT about something that I see the BBC have labelled ‘tokenomics’1 (hats off to that subeditor) – how on earth one begins to measure the value of the tokens which companies of all stripes are now budgeting for their employees’ use of AI agents. In his role, he get 300 tokens a day, and every now and again he maxes that. But, as the BBC explain well, there’s no real way to anticipate how many tokens a particular task is going to consume, so budgeting your daily work is tricky at best.

I continue to have a stack of AI-related articles accumulating in my inbox – as you may well also, if you have subscribed to really anything at all in 2026.

I have The Gospel Coalition telling me they’re banning the use of AI in writing articles submitted for publication2 – a line many publications are drawing, though most won’t publish a whole manifesto to that end.

I have Cal Newport gently dismantling the recent publicity-generating stunts around AI agents ‘escaping’ from their sandboxes – an actual expert pouring quite a lot of water on the fire, whilst also drawing attention to some real onrushing issues.3

I have Jon Gruber highlighting this excellent and revealing essay by Nikhil Suresh, where Suresh details what many thinking people suspect – even if big businesses have no use for AI in a particular area, they’re shoehorning it into every possible nook or cranny, because their clients assume it represents some kind of whiggish progress – a sort of, ’10x EVERYTHING’ mentality, even if we’re already seeing some evidence that adding AI agents does not streamline work – it just changes it to new work.4

And I have (via Jason Kottke), David Reichert’s deep dive into a distinction I think we’re struggling to delineate right now, but which is rapidly developing – the gap between what most people think of as AI – which are really free bits of software we can use to generate language responses or images, or even video clips – and the hugely powerful agents which have already transformed programming in business.5

How do we reflect on all this? Well for one thing, it strikes me we need our descriptors to evolve, and some new nouns to be added.

I’ve spoken a lot, and written a little, about the limitations of LLMs, but nevertheless Joe Public will continue to call ChatGPT or Gemini ‘AI’. But if you speak to a coder, they’re using the term to refer to hugely powerful models that they are spending their days speaking to through terminals and deploying for work purposes.

A little randomly, I was on GitHub the other day, looking at the account of someone I know, and noticed that in the last year his daily commits across a range of projects (he leads a design company) have gone from a handful to dozens a day. But of course they have: he’s now able to deploy models to write, test, and iteratively improve what he’s working on at a pace we could only dream of in the days of tapping out every line of code by hand.

We need new names for the models consumers are using, versus the actual revolution that’s happening in the backend of software. I don’t much trust either, although they absolutely have their uses – but I wouldn’t mind the language to talk about them more clearly.

  1. Joe Fay, ‘Tokenomics: Why making AI pay is tricky‘, BBC News, 12 August 2026. ↩︎
  2. Brett McCracken, ‘Our Line in the Sand on AI: TGC’s Missional Opportunity‘, The Gospel Coalition, 5 August 2026. ↩︎
  3. Cal Newport, ‘Did OpenAI’s New Model “Go Rogue”?‘, CalNewport.com, 27 July 2026. ↩︎
  4. Jon Gruber, ‘AI Mania is Eviscerating Global Decision-Making‘, Daring Fireball, 25 July 2026. See also: Slack; email; every other transformative technology that just ended up generating more administration and less deep work. ↩︎
  5. David P. Reichert, ‘If you haven’t recently used Claude Code, you might not understand where AI is at‘, Substack [Yuck], 3 August 2026. ↩︎