Seroter's Daily Reading — #870 (September 18, 2026) · Seroter's Daily Reading

Listen: https://blossom.buildtall.systems/c5f81d5a04b977112471f11203a5338cfce2aafb719bce23e3fb59dea38d7b4c.mp3

Episode 870 of Seroter's Daily Reading, for September 18th, 2026.

Let's start with an idea that's been circling the developer world for a while and finally got spelled out clearly: We are all managers now. InfoWorld ran a piece arguing that coding agents aren't perfect and that, like human employees, they need clear instructions, wise guidance, and the right working environment. The key line comes from Jim Bourke of Zide Software: a coding agent is most useful when you treat it more like an employee than an autocomplete tool. It needs a clear assignment, the right context, a safe place to work, checkpoints, and a way to evaluate the result. Managing agents isn't the same as managing humans, of course, but the list of what an agent needs does look awfully familiar.

This week also brought a genuinely refreshing take on obsession, from Steve Magness over at The Growth Equation. His essay is titled Stop Trying to Be Less Obsessed, and it's built on the research of psychologist Robert Vallerand, who splits passion into two kinds. There's obsessive passion, where your self-worth is on the line and you feel compelled by pressure, and harmonious passion, where you genuinely choose the work and it fits your life. The piece is remarkably honest about the fact that obsessive passion is powerful rocket fuel, but the downsides compound as you age. The real advice isn't to obsess less. It's to build up the rest of your life so the obsession isn't the only thing holding you up. That's a useful framing for anyone who's ever been told to just chill out.

There's a big infrastructure story this week too. Zed launched the public beta of Delta, and the headline framing is that everyone is in a race to replace GitHub. The reasoning is that pull requests, a fifteen-plus-year-old interaction model, don't fit a world where agents generate enormous volumes of code and the real reasoning lives inside a conversation rather than a diff. Delta is built around the thread, a running record of an agent-assisted coding task where the conversation and the files stay connected, with a new operation-based version control layer called DeltaDB underneath. Zed has already disabled pull requests on Delta's own repository and landed more than 570 changes without them. I'll admit I'm still skeptical that GitHub itself is in peril, but challenging assumptions behind classic Git is healthy, and Zed is far from alone here. Cursor and GitLab are pushing in similar directions.

On the same theme, James Governor at RedMonk wrote about agents choosing software infrastructure now, and how we're not really ready for it. He connects it to Stephen O'Grady's argument that agents are the new, new kingmakers. Agents aren't just writing code. They're often choosing the language, the libraries, and the frameworks, sometimes even against explicit instructions. The most striking data point James cites is a chart from Render showing a sharp jump in paid services created starting in January 2026. His answer to what happened that month is simple: capable autonomous agents happened. The human in the loop started to feel anachronistic. The takeaway for all of us is that agent choices are production choices now, whether we've fully internalized that or not.

Switching to consumers, Menlo Ventures published its 2026 State of Consumer AI report, and the headline is fascinating. Adoption barely grew, rising just three points from sixty-one to sixty-four percent of US adults, but spending tripled, from twelve billion dollars to roughly forty billion. So the story isn't new users. It's existing users going deeper. Fifty-five percent of AI users now pay for at least one AI product, and the top fourteen percent of payers account for sixty percent of spending. Nearly half of AI users have earned money with AI, mostly through content creation and freelancing, which translates to about thirty percent of all Americans. And trust is becoming the real gatekeeper, with accuracy, trustworthiness, and privacy now ranking ahead of ease of use. It's a market maturing fast.

For the builders out there, there was a practical piece on scoping a Flutter MVP so founders get a first-value moment in weeks. The core idea is that fuzzy scope is where founders quietly lose weeks, so you define a first-value moment, the earliest measurable user action that proves the product delivered its promise, and build strictly toward it. The article lays out some useful disciplines: lock a thin design system early, keep auth and backend boring on purpose, and run every feature through three questions about whether it's truly needed to test the hypothesis. There's even a founder-versus-studio ownership checklist and a realistic six-to-ten-week timeline. It's pragmatic advice for anyone shipping a mobile product in 2026.

The Harvard Business Review asked a pointed question this week. You think you're a good collaborator, but do the people you manage agree? The piece opens with the observation that most senior leaders believe they're effective collaborators who emphasize listening and empowerment, yet the data from the people they actually interact with tells a different story. It's a worthwhile gut check for anyone leading a team, and the suggestions for closing that gap are worth sitting with.

There was also a sharp essay arguing that AI is making activity-based engineering metrics obsolete. The evidence here is a scatter plot, not a coefficient. A METR randomized trial on expert maintainers showed AI made them nineteen percent slower, while enterprise field experiments showed twenty-six percent more completed tasks, and observational data on heavy Copilot users suggested up to forty percent more pull requests. The point isn't that any single study is wrong. It's that metrics like lines of code, commit count, and story points can no longer tell you which situation you're in. The recommendation is to demote the activity proxies and track review queue time and short-horizon rework instead, always trended against a pre-AI baseline. And Goodhart's Law looms over the whole thing, so those dashboards should never become individual leaderboards.

Finally, Google published a post on why client SDK generation belongs in the open, and the catalyst was blunt. In May 2026, the SDK generation provider Google relied on was acquired and abruptly shut down, right as they were gearing up for the Interactions API at Google I/O. That disruption showed that proprietary, closed-source generators create unacceptable platform risk. Google partnered with Speakeasy to migrate in place and help open source the full generator suite under AGPL. The result generates multi-language SDKs, agent-native CLIs, and even documentation MCP servers, so agents can query live schemas instead of guessing at outdated methods. It's a concrete example of infrastructure betting on openness, which honestly ties back to the agent theme running through much of today's list.

If there's a throughline today, it's this: agents are no longer a novelty. Whether we're managing them like employees, rebuilding version control around them, watching them pick our infrastructure, or noticing they've begun earning us money, they've moved from the edges to the center of how software gets built and used. That's it for episode 870. Thanks for reading, and enjoy your weekend.

Your signing session ended