Agents are changing what software is. Understanding is what keeps up.
The shift isn't that we type code faster. It's that agents are starting to do the work, and software is turning into something you describe instead of something you write.
I think that's a paradigm shift. The kinds of tasks we're focusing on are just different. What took a lot of people a lot of time is shorter now, and we're going to be touching new problems.
The clearest picture of it I've found is a podcast in October 2026 with Mark Zuckerberg and Alexandr Wang, launching Muse Wonder, the new model behind Meta's AI assistant, Muse. They're selling it, so take this as their view. But it's a view from the people building the frontier.
What they're describing
Everyone writes code. Zuckerberg asked his AI to coach his martial arts training between rounds. It built itself a video pipeline, pulling frames out of the footage. He never saw the code. He steered it by results only: it took 30 seconds to give feedback, his break is 60 seconds, so it had to go faster. His prediction is that lots of people will write lots of code. "They're just not going to know that they're writing code." (0:02:02)
The model is the product. Zuckerberg says the line inside Meta is that the model is the product. The personality and the capabilities get trained into it. A lot of the product work around it is privacy and security. (0:12:01) Wang adds that even product decisions now come down to training methods, data and evaluations. (0:13:00)
Small teams win. Zuckerberg recalls the first advice Wang sent him for rebuilding Meta's AI effort: you don't need a lot of people. Talent density is what matters. (0:32:01) Zuckerberg says the other half is giving each of those people a lot of compute, instead of spreading it across competing efforts. (0:33:01) By his own account the research team is under 200 people. (0:34:02)
It's moving fast. Wang thinks AI is maybe five or ten times faster moving than the tech waves before it. (2:05:01) He also points out that really small models today are much more powerful than the biggest models from a year or two ago. (1:39:00)
And it's about making things. Zuckerberg says their belief is that the highest use of superintelligence is invention and creation, not automation. (1:07:01) It's also his pitch against rivals focused on automating office work. But it's the version of the future I'm optimistic about.
It's changing commerce too
Zuckerberg used Instacart maybe twice a year. Once his AI assistant could come up with recipes with him and order the groceries, he started using it once or twice a week. He's now a regular customer only because the agent took the effort away. His point is that the opportunity isn't making existing tasks 20% more efficient. It's new activity: people doing things they wouldn't have bothered with before, and businesses getting customers they wouldn't otherwise have had. (1:16:02)
Their bigger claim is that this changes how buying works. Zuckerberg says that just as every business today has a website, social media accounts and an email address, every business is going to have agents too. He also says far more businesses are calling Meta to integrate than trying to block it. (1:09:01)
Wang describes Muse as one of the first products where people fully hand a purchase to an agent. You say something like "I need something by Wednesday", and the agent does the research, finds the product and fills out the order. Next to that, keyword search starts to feel primitive. (1:11:02, 1:12:01)
Wang expects shopping through agents to grow as a share of all shopping, the way online shopping grew alongside physical stores. (1:10:03)
The host, John Collison, raises the sharp edge. Businesses that only work because people spend time on their site, or because they overcharge, have a reason to block agents. Zuckerberg's answer: those are the ones that get disrupted. (1:08:01)
Wang says he saw it coming. In 2023 he told Stripe's Patrick Collison, John's brother, that agents were going to be buying things and Stripe wasn't ready. In his words, he was right about the direction, but about three years early. (1:19:00)
Where the money is
Zuckerberg is direct that this is about money, for people and for Meta.
Saving it. His example is Muse going through your email, finding the subscriptions you aren't using, and cancelling them. (1:28:03)
Making it. For small businesses, he wants Muse to be a "business in a box": an agent that helps run the admin side of the business. (1:29:00)
Meta's share. He says Meta could, over time, take a small cut of what Muse saves or makes for people, and he expects that to become a very large revenue stream. On top of that, Meta charges subscriptions for the bigger model, sells it to businesses through an API, and has its existing business. (1:28:03)
So the people telling the money story also plan to take a share of it. That's worth knowing when you hear it.
Where the gap shows up
All of that raises a question. If people are handing over the work without ever seeing it, who understands it?
Collison asks a version of it himself. When the model says Zuckerberg is too far on his back leg, does it understand balance, or has it just seen a lot of pictures? Zuckerberg's answer is only that they trained it on a lot of visual data. (0:04:03)
That's the understanding gap. The distance between something that works and someone who knows why it works. Where it breaks. Whether it's the right thing to build at all.
It isn't only code. Wang describes students talking stream-of-consciousness for 10 or 15 minutes, then having AI organise and structure it for them. (1:23:03) That's fast. It's also the structuring, the thinking part, being handed over.
The builders know this. Wang says a lot of the magic is that it just works, and if it were even a little less reliable, people would throw the app away. Zuckerberg credits that to a core group of people on the product team who focused on getting it right. (1:18:02)
What the builders protect
Here's the part I find most interesting. The people building the agents are the strictest about understanding.
Zuckerberg says the point of Meta's lab is the absence of non-technical managers. People who step out of the details lose their understanding faster than they realise, so people have to stay technical, use the agents, and write code. (2:06:01)
They protect it in three ways.
At the top. Soon after Wang joined, one of the leads asked whether consumers even needed a reasoning model. Wang calls it a heart-sinking moment. His takeaway was that Meta's leadership had to get far more "AI-pilled", and he says it did. Instagram's head, Adam Mosseri, now has one of the most effective personal setups of vibe-coded mini apps Wang has seen. (2:03:05, 2:04:01) The leaders understand it because they use it.
In what they measure. Wang says steering by public benchmarks means outsourcing your product thinking to whoever happened to make them. Meta's own evals are a long spreadsheet of small details, each maybe worth 2% to a user. All of them together are what make it feel magical. (0:29:02) You can't write that spreadsheet without understanding the product.
In judgement. Even at the frontier, it isn't all numbers. Wang says researchers squint at charts to decide if a model is good, and some hold physical protractors up to their monitors to measure a curve. (1:53:00) Someone still has to know what a good curve looks like.
The shift
Agents are going to do more of the work. That's the paradigm shift, and I'm optimistic about it.
There's a catch Zuckerberg admits too: probably only a small handful of companies will be able to build this at the frontier. (1:06:03) But the people building them aren't handing over the understanding. They're holding on to it tighter.
The episode ends on the same note. Collison says it's easy to fall behind, and you have to stay at the frontier. Zuckerberg's last word on it: stay in the details, stay technical. (2:13:01)
If you love learning, this is a great time to close the gap.
Source
- Cheeky Pint, Muse Wonder: Zuckerberg and Wang's $100B AI, 12 October 2026.