Seroter's Daily Reading — #867 (September 15, 2026) · Seroter's Daily Reading

Listen: https://blossom.buildtall.systems/cb79527440054d728df8918949aec1b09f59664188a545ddd6d970dd3d3f7e31.mp3

Welcome to Seroter's Daily Reading, episode 867, for Tuesday, September 15th, 2026. I'm coming to you from Boston, where I'm getting ready to keynote a conference tomorrow. Before I hit the stage, I pulled together a handful of items on modern developer tooling, voice AI apps, and how we should be thinking about software development right now. Let's get into it.

First up, a piece that's really about a problem we don't talk about enough: workspace coordination. Steve Yegge posted a question on X asking how people manage ten or twenty coding agents at once, because, as he put it, he uses Emacs and wouldn't wish it on anyone. The replies, six hundred and counting, offered everything from Herdr and cmux to Conductor and the Claude and Codex desktop apps, plus a pile of homegrown tools. But the interesting part isn't which tool people picked. It's what developers have had to bolt on to these tools to make up for their shortcomings: shortcuts to find the agent that's waiting on a decision, ways to recover an earlier conversation, and groupings that show how a task fits into a bigger project. So even with all this incredible intelligence available, we're still struggling with something surprisingly simple: remembering what we asked the AI to do. There's no settled answer here yet, and that's worth paying attention to.

Next, Google is pushing us to get out of the chat box and into real-time voice. The company released Gemini 3.5 Transcribe for low-latency speech-to-text, with a four percent word error rate and some genuinely useful features, like automatic code-switching across languages, custom vocabulary biasing so it can handle your domain jargon and company names, and a smart transcription mode that cleans up filler words and disfluencies. Combined with Gemini 3.8 Live, the pitch is real-time voice applications with sophisticated tool calling. The point Seroter is making is that we should think about how this changes the way you and your customers get things done, not just chatbots but actual spoken, conversational workflows.

Sticking with Google for a moment, Thomas Kurian gave a keynote at the Goldman Sachs conference, and he's very good at boiling things down to three points. His pitch: Google Cloud is the only provider spanning the entire AI stack, which expands their market and differentiates them on performance, cost, and quality. They've got seventeen product lines with more than a billion dollars in revenue each, and they've seen more than double the number and value of hundred-million-to-billion-dollar deals year over year. He also talked up the AI infrastructure story, with TPUs paying back faster than GPUs, and noted that customers who use their AI products end up using one point eight times as many products overall. It's a confident, numbers-heavy case for the full-stack approach.

Now for one of my favorite items today, a genuinely useful post on pair programming with your AI agent. The author argues that when you open an agent in your project directory, you're already pairing: shared context, shared goal, one place the code lands, and a conversation running alongside. The only difference is your pair reads faster and never needs coffee. He maps the four classic pairing styles, unstructured, driver and navigator, strong-style, and ping pong, onto working with an agent, and the advice is practical. For example, in strong-style pairing, you never touch the keyboard; you state intent, then location, then detail. The key insight is that you can just tell the agent which style you want, because that vocabulary is already in its training data. My favorite takeaway, though, is this: pairing gives you comprehension, but only running the whole thing gives you confidence. So keep a way to execute your full environment, even if it's a throwaway test harness.

Related to that, BigQuery now has a set of augmented analytics functions that go way past basic queries. These are table-valued functions for things like key drivers, causal effect, correlation, change point detection, trend, and seasonality. The cool part is that you can chain them together, and because they output structured SQL, they can be plugged in as skills for AI agents, enabling conversational data investigation. So instead of just asking an AI to run a query, you can ask why a metric moved and actually get causation and trend analysis. Seroter's point is that we're well past simple queries now.

Then a lighter one. Harvard Business Review has a piece on leaders and humor, which made Seroter sigh because looking at the science of humor feels weird, but the advice apparently makes sense. The research shows humor can grab attention and build connection, but a failed joke diminishes a leader's social capital, lowering respect, likability, and perceptions of competence. So the question is why some jokes land and others fall flat, and how leaders can do better. Whether Seroter will actually follow the advice, well, nobody knows.

The longest and most thought-provoking read today is a post arguing that we are all product engineers now. The author's thesis is that the cost of actually writing code has collapsed, and it's going to keep collapsing through review, maintenance, and eventually shipping and scaling. What's left, and what's durable, is deciding what to build, defining what good looks like, and making it delightful. The junior developer job, which was mostly writing well-specified tickets, is essentially gone, and the new job that's emerging, under names like forward deployed engineer, is someone who sits with the customer, figures out what they actually need, and builds it with the agent doing most of the coding. Those roles are paying a quarter million dollars and up. It's a sober, realistic forecast, and Seroter found little to disagree with.

Finally, a shorter one on agent permissions in Antigravity, the point being that YOLO mode is fine for demos, but even if you trust your agent, you'll eventually want a more granular, thoughtful approach to what it's allowed to do. The post itself was behind a security wall when it was pulled, but the principle stands: as agents get more capable and more autonomous, permissioning them carefully is going to matter.

So that's the through-line today. The tools are getting smarter, but the interesting problems are shifting from writing code to coordinating agents, deciding what to build, and setting the right guardrails. Whether it's workspace coordination, voice interfaces, or the future of the developer role itself, the human part of the job is becoming the product thinking and the judgment. Thanks for reading along, and I'll talk to you next time.

Your signing session ended