Seroter's Daily Reading — #854 (August 26, 2026) · Seroter's Daily Reading

Listen: https://blossom.buildtall.systems/89f56f7077df5f2dee4c406225c5ec10e29c38004a25fead7d213fe060ec6dfe.mp3

Welcome to Seroter's Daily Reading, episode 854, for Wednesday, August 26th, 2026. Today's batch is a good one, and it touches on a few themes that keep coming up: how you train your people, how you share knowledge across teams, and how you pay for AI services. Let's dig in.

First up, a piece from Google Cloud asking what harness engineering is and why you should care. If you've heard the term thrown around and nodded along without really knowing, you're not alone, and this is a good chance to catch up. The short version is that harness engineering is about the scaffolding around a model, the code and tooling that turns raw intelligence into something reliable and useful in production. It's the difference between a clever demo and a system you can actually trust. Even if you think you know the space, there's usually something new to learn here.

Next, a really useful framing from RedMonk on the difference between Generative Engine Optimization, or GEO, and Agent Experience, or AX. The one-liner is memorable: GEO gets you mentioned, and AX determines what happens next. GEO is the new SEO, the work of making sure your product shows up when a model answers a question. AX is everything after that, whether an agent can actually get in, find what it needs, and finish a task inside your product. The piece also has a nice reality check on llms.txt, the proposed file that's supposed to point models at your best pages. Turns out it's mostly being fetched by crawlers, not by the agents it was written for. So if you sell to people, GEO is your problem. If you sell something an agent has to operate, AX is your problem.

Then there's a thought-provoking piece on whether the future of learning is what the author calls vibe training. The core idea is that training has spent a century focused on the how, the step-by-step process, and almost no time on the why. But in a world where the how changes constantly and models can handle a lot of it, the why is the thing worth teaching. The author borrows from TikTok: short, snackable, one or two concepts max, with a good hook. And he suggests generating training on the fly, at the moment someone needs it, rather than shipping a fixed curriculum. It's a shift that matters as much for trainers as for students, because it forces you to be clear about what actually matters.

Now a couple of pieces on the money side. Stripe published five monetization trends from pricing leaders, and the headline is that the old revenue playbook is breaking down. Pricing is becoming always-on, with companies like Lovable making ten pricing changes in a single year. The obsession with annual recurring revenue is holding companies back from usage-based models, even though top-ups often behave like recurring revenue anyway. And the agent customer is arriving, which means pricing and payments need to be redesigned for a buyer that isn't a person. Vercel's CEO talked about sweating the details of error messages, with the agent as the customer.

Google Cloud followed that up with a FinOps-for-AI announcement: new flexible billing and cost controls for agent workloads. The idea is to mix predictable per-user seats with pay-as-you-go, so agent workloads don't hit quota limits mid-task. There are flexible savings plans that give you ten to twenty percent off token costs for a spend commitment, and there's a coming-soon deferred execution option that runs eligible workloads during off-peak windows for up to half the inference cost. The through-line is the same as Stripe's: how we consume these products is changing, and the billing has to change with it.

PostHog has a candid piece arguing that you need to find product-market fit again, sorry. The point is that the product-market fit game never really ends, especially now that a two-person team can build your app for 2026. The author lays out a playbook for disrupting yourself: write down how you die, fund the attacker yourself with a small team, score it like a seed bet rather than a business unit, and build first and validate later. One detail I liked: they ran the classic how disappointed would you be survey on their own team before a launch and scored twenty-three percent, well below the forty percent benchmark. Your own team will tell you the truth before your churn does, but only if you ask.

Daniela Petruzalek wrote up building agents in Go with Gemini, and it's the test I wanted to run, done better than I would have. It's a solid assessment of the different ways to assemble agents in Go, and worth a read if you're in that ecosystem.

There's also a slide deck from a16z called Intelligence is the Primitive, Applications are the Diffusion Layer. It's a set of slides that challenged some of my own thinking, particularly around moats, or the lack of them. The framing is that raw intelligence is becoming the underlying primitive, and applications are just the layer that diffuses it out to users. If you're thinking about where durable value actually lives, it's worth sitting with.

Flutter published a fun behind-the-scenes look at how they stay ahead of iOS releases. Every June, Apple drops over a hundred technical sessions at WWDC, and this year the Flutter team started using Gemini to triage all of them, ranking each session by importance and impact so they know what to actually watch. It's a nice example of using AI to make sense of a firehose of platform change, and it's how they keep shipping day-zero support for every major iOS release.

Then a genuinely interesting research question: what language are agent skills written in? Skills, those SKILL.md files that give an agent instructions in plain prose, are spreading fast, millions of them across hundreds of thousands of repos. And the data shows that in early 2026, thirteen percent of newly written skills were in a language other than English, jumping to sixteen percent a quarter later. That's a three-point move in three months, and it's a floor, not a ceiling, because English isn't a proxy for American. The point is that AI development has arrived somewhere other than San Francisco, and it's happening fast.

The big news of the day is Salesforce putting its entire CRM inside Claude. The partnership, called Claudeforce, ships a plugin with thirty-seven pre-built sales skills that let sellers query, update, and act on live CRM data without ever opening Salesforce. It's built on the headless approach Salesforce started earlier this year, exposing APIs and MCP servers so agents can call the platform directly. The striking part is the framing: Salesforce spent twenty-seven years building the defining interface of enterprise software, and now it's telling customers they may not need that interface at all. The value, they argue, was never the screens, it's the data and the workflows. That's a harbinger of where a lot of software is heading: build more endpoints, and fewer frontends.

Finally, a practical piece from Google Cloud on using the Open Knowledge Format with Knowledge Catalog to serve context to agents. OKF is a portable format for the context agents need, and the question this post answers is how you share and govern those bundles across an organization. The answer is to map them onto a knowledge catalog, so every concept becomes discoverable, governed, and reachable by any agent already reading from it. I actually got asked in a meeting today how to distribute shared context, and I had no good answer until I read this.

And that's the thread running through today's list. Whether it's training, knowledge, or pricing, the underlying shift is the same: the how is being automated, and the durable work is moving to the why, the context, and the trust. Thanks for listening, and I'll see you next time.

Your signing session ended