Millie: Right, so get this: in your average big company now, the number of machine identities outnumbers human users forty-five to one. Forty-five to one!
Pierre: Forty-five... that is significant. So, more machines accessing systems than people? Already?
Millie: Yeah! It's like, I read that and my brain just went 'whoa, okay, this is a different world.' And it got me thinking about this article, about how software is changing because of this. It's not just adding a chatbot to your website, you know? It's deeper.
Pierre: Yes, I was seeing something similar. The article mentioned Neon, the database company, saying eighty percent of their new databases are created by AI agents. Not by people clicking buttons, or typing commands themselves. And GitHub, over five percent of all code commits are completely authored by AI.
Millie: Five percent of all commits? That's not just helping, that's... doing the job.
Pierre: Exactly. And this is the core idea, no? That the 'user' for software has changed. It's no longer just you and me sitting at a screen, clicking. It is these agents. The product's value is not its beautiful interface, but how well it performs a task for a machine.
Millie: Yeah, that's what I was getting at! It's not AI assisting human users, it's AI being the user. The software needs to be rebuilt for them, not for us. It's a proper paradigm shift, isn't it?
Pierre: A complete shift. And for me, this feels a bit like the early days of AWS, you know? You could go into the console and click around, provision a server, set up a database. But all the real power users, all the automation, it was all happening through the command line interface, the CLI. The GUI was just for, perhaps, getting started, or a quick check.
Millie: Oh, that's such a good connection! Like, the 'real' work, the scalable stuff, it was never through the pretty picture. It was always under the hood, programmatic.
Pierre: Precisely. And we are cycling back to that. The GUI is for the one-off tasks, maybe the onboarding, but the product itself, the real value, is the API, the command line. So, how do companies actually build for this? There are three big things the article points out.
Millie: Right, I'm keen to hear this, because it feels a bit... abstract, thinking about software for machines.
Pierre: First, it's about 'skill files.' This is where a company's domain expertise, the 'how-to' knowledge, becomes machine-readable. PostHog, they learned this. They had to write a specific 'skill file' for their AI agent. Why? Because the default way an agent might calculate user retention, for instance, by just looking at 'signed_in' events, would give misleading data. You need to tell it, no, use this specific, non-obvious event instead.
Millie: So it's like, writing down the unwritten rules of the business, but for an AI. All that tacit knowledge, the stuff you learn after years on the job, now it has to be codified.
Pierre: Exactly. That deep practitioner knowledge, living in a markdown file for an agent. The second thing is CLI tools. The command line, it's the new user interface for these agents. Take 37signals, they rebuilt Basecamp with a full-featured CLI. Their CEO, DHH, he was quite blunt: 'Agents have emerged as the killer app for AI... we’re launching a fully agent-accessible version today.' They made their entire product usable by a machine.
Millie: That's wild, going back to the command line for a product like Basecamp. I mean, my brain still goes to, like, a spreadsheet when I think of that kind of tool. Not a blinking cursor in a terminal.
Pierre: But for an agent, a command with structured input and output is far more composable, more reliable, than clicking through a complex GUI. And finally, vertical models. This is about building or using specialized AI models that are cheaper and better for specific domains.
Millie: So, like, not the big general models, but little, focused ones?
Pierre: Yes. Cursor, the AI code editor, they launched their own coding model. And it actually beats Claude Opus on certain benchmarks, but at one-tenth the price. They use the big, expensive frontier models for the absolute hardest problems, but for everything else, they use their own cheaper, specialized model.
Millie: Okay, that makes sense. Like having a specialist for a specialist job. But then, it also mentioned Harvey, the legal AI, that built its own model and then... scrapped it because the bigger models caught up? So are vertical models a good idea or not?
Pierre: Ah, that is where it gets complicated, n'est-ce pas? Harvey's specific legal model, it was superb for a time, lawyers preferred it. But the baseline, the general models, they improved so fast, their unique advantage faded. It shows that unless your domain is truly unique, or you have a massive proprietary data advantage, relying purely on a vertical model might be a tricky long-term strategy. You need to keep investing to stay ahead of the general models.
Millie: Right. So it's not as clear-cut as 'build your own model.' You have to really know if your special sauce is special enough, and for how long. It's a lot to think about. Okay, let's switch gears a bit.
Pierre: Oui. Because there is also this question: is this agent-centric future just... a developer's fantasy? Most business users, they are not engineers. They don't want to manage armies of agents via a command line, or write markdown skill files. They just want a smarter GUI that does the work for them, with all the agent complexity hidden under the hood.
Millie: That's exactly what I was thinking! My mum isn't going to fire up a terminal to book her flights, is she? She just wants the website to be smarter, to anticipate what she needs. This whole CLI thing, it feels like it's going backwards for the average person.
Pierre: But that is the point. The 'average person' is not the primary user anymore for these new software paradigms. It is the agent. And the agent prefers the CLI. It is more efficient for them.
Millie: But isn't there a risk that if everything becomes a CLI or an API for an agent, then the actual product itself gets commoditised? If an agent can use any tool, then all the value just accrues to the agent provider, not the tool maker. You just become a cog in someone else's machine.
Pierre: Ah, you mean the agent becomes the 'app store,' dictating terms? It is a possibility, yes. But if your tool is truly superior, even programmatically, the agents will still choose it. Performance, remember? That is the new competitive advantage.
Millie: Yeah, but 'superior' for a machine might just be 'cheaper' or 'faster' in a way that doesn't feel like a unique selling point for a human business owner. It's a tricky balance. I'm Millie.
Pierre: And I am Pierre. This has been Manish Chiniwalar's Station.