Seroter's Daily Reading — #860 (September 3, 2026) · Seroter's Daily Reading

Listen: https://blossom.buildtall.systems/cad4b1c3a4297017ad215d4650cf0a973bcb30d7f4e06d4bf8d510d10e1ed097.mp3

Episode 860 for Thursday, September 3rd, 2026.

Today's list is a genuinely wide mix. There's some practical advice, some geeky deep dives, and a healthy dose of hard truth. Let's get into it.

We start with one of the more provocative pieces I've read lately, from Latent Space, called PRs NOT Welcome. GitHub invented the pull request, and for eighteen years it's been open by default. But now a bunch of top AI-native open source projects are shutting PRs off entirely. The reason is blunt: a lot of incoming pull requests are AI-generated, and maintainers would rather use their own agents than triage a flood of community code. Vercel wrote up its software factory for the AI SDK, which gets over twenty million npm downloads a week and had a backlog of over a thousand open issues and almost eight hundred pull requests. They deployed specialized agents, one to reproduce a bug, another to apply a fix, another to review it. Four weeks in, those agents author between twenty-five and thirty-five percent of the PRs they merge and close seventy to eighty percent of issues. Astro's creator Fred Schott adopted the same idea, and it inspired a whole new agent framework called Flue, which simply closes every external PR and turns it into an issue or discussion. The tldraw folks do the same, and Mitchell Hashimoto thinks large open source projects will eventually close contributions completely. The interesting tension is community. Pull requests were historically how new contributors got mentored and became maintainers. If agents do all the work, the remaining door is discussion, reporting issues, and building trust through conversation. It's a big shift in what open source collaboration even means.

Next, a quick pair of Google Cloud notes that had their full text blocked, so I'll go off the summaries. One covers taking advantage of Gemini Managed Agents with Google Apps Script: people build wild automations in Apps Script, but there are real limits, so calling out to Gemini Managed Agents becomes a powerful pattern once you outgrow what a script can do alone.

Related, a series is kicking off on the three identity kinds on Google Cloud: human, machine, and agent. Seroter's question is a good one, why do we need all these different identities and what's each one for.

On the vibe coding front, InfoWorld has a piece on seven critical vibe coding mistakes and how to avoid them. Seroter's skeptical take: are these really the worst mistakes? Six of them don't even apply to a demo app. But the one that stands out is dependencies. The argument from Sonatype is that AI isn't just writing your code, it's also deciding which open source components you depend on, and if those choices aren't informed by current intelligence, you can quietly build on outdated, abandoned, or risky libraries. The fix is to specify a catalog of approved components, keep it fresh, and review the software bill of materials on everything your vibe coding tools produce.

Over at Google DeepMind, they've introduced WeatherNext 3, a genuinely impressive leap. The big difference from previous AI weather models is what it learns from. Most train on numerical weather prediction output, which comes with a six-hour data lag. WeatherNext 3 ingests live satellite observations instead, generating a fresh forecast every hour at up to five kilometer resolution. That matters because storms develop fast, and conditions can swing wildly over just a few kilometers near coastlines, valleys, or mountains. It also forecasts the hundred-meter wind speeds and cloud cover that renewable energy operators need to predict output, and on precipitation they report up to a sixty percent improvement in forecast skill against one baseline. Seroter's note made me smile: if this saves even one person from getting rained on after a confident clear skies forecast, it's worth it.

There's a piece called Software Is Fast Fashion, and the full text was blocked, but the premise is worth chewing on. The analogy is that we treat software as if it's built to last forever, when a lot of it should really be treated as disposable. Seroter leans toward some agreement here, that there's far more we should be treating as throwaway than durable.

Battery Ventures has a long, rich essay called Defying Gravity: The System of Record in the Age of AI, and it's one of the more thoughtful pieces today. It dismantles the narrative that the system of record is dead by focusing on market turnover. Big systems like CRM and ERP are huge markets, but they have high retention, so only a tiny fraction of that spending changes hands each year. The essay shows the math: a five billion dollar market with ninety percent gross retention leaves only about five hundred million in jump ball ARR a year, and after you account for the deals you see and win, you might grow at twenty percent at five hundred million in ARR. But AI changes the equation on both ends. First, the desire to adopt AI is breaking those locked markets open, because you can't point an agent at fragmented, clunky data and expect it to work. Second, the total addressable market is growing, because a system of record that owns the data can now automate the actual work, shifting from selling seats to selling outcomes. The result is businesses that can sustain eighty percent growth at a billion in ARR. The conclusion is blunt: the window to build a new system of record is open right now, because customers are AI curious and their existing vendors aren't serving them, but it won't last forever.

That optimistic essay sits in tension with the next piece. According to Gartner data from CIO Dive, fewer than a quarter of enterprises, twenty-two percent, have successfully scaled AI across multiple business units. But the lag isn't slowing anyone down: eighty-five percent of tech leaders plan to increase AI investment next year, even as eleven percent admit they have no visibility into what they already spent. The bigger stat: nearly three quarters of executives scaled fewer than a quarter of their AI pilots successfully. The through line is measurement. Companies that track ROI constantly, treat AI as a portfolio, and kill underperforming projects reported positive returns in eighty-one percent of initiatives. If you don't know what you're aiming for, it's hard to call anything a success.

Pinterest Engineering published Becoming an AI Team, though the full text was blocked here. Seroter's takeaway is that it covers how their engineering org evolved into an AI-enabled team, with a lot of breadth and good visuals. If you're planning your own transformation, it's probably worth a look.

Google Cloud was named a leader in the 2026 Gartner Magic Quadrant for Strategic Cloud Platform Services, the first new entrant to the leaders quadrant in years. The post hangs the recognition on three things: a co-designed stack from silicon like TPUs and Axion up through GKE and Gemini; a dynamic infrastructure meant to absorb the bursty nature of scaling AI; and sovereign cloud options including air-gapped environments. One stat jumps out: Google's data centers now process 3.2 quadrillion tokens a month, roughly seven times more than last year.

OpenAI posted a quasi-release today called Path to Astra, though the page just served a Cloudflare challenge when it was fetched. Seroter's note says it's not readily available until next week, but Astra looks like a monster. I'll keep an eye out for the details once it's actually live.

For the Go developers out there, the Go team published a deep technical dive on goroutine leak profiles, a new feature in Go 1.27. A goroutine is leaked when it's blocked on something that can never unblock it, and those leaks build up silently, eating memory and pinning the garbage collector. The new profiler is notable because it's precise, with almost no false positives, unlike older goroutine profiles that can't tell a leak from a traffic spike. The clever part is how it works under the hood: it borrows the garbage collector's reachability machinery to figure out which blocked goroutines are unreachable by any live goroutine, then reports those as leaked. There are limits, it only covers channels and sync primitives, not file or network IO, and it can't catch non-deterministic leaks before they happen, but for a huge class of concurrency bugs it's a welcome addition.

Rounding out the list, there's a piece from Prashanth on taking your AI harness with you via Antigravity Remote Control, letting you control your coding bot from anywhere.

There's also an HBR piece called Middle Managers Will Make or Break AI Adoption. The HBR argument is that the most common failure point in generative AI isn't the boardroom or vendor selection, it's the layer of managers who have to turn executive ambition into everyday behavior. Seroter's version is sharper: if you're not focused on the right internal population, you'll either get your middle managers to champion the new way of working, or watch it all go up in smoke.

That's the thread that ties today's list together. On one hand we've got weather models and software factories and goroutine profilers running fast and doing remarkable things. On the other, we've got the very human reality that most enterprises still haven't scaled AI, and the people problem in the middle is often what determines whether any of it lands. The tools race ahead; the organization is usually the slower work. That's it for episode 860. See you next time.

Your signing session ended