Agentic Engineering Weekly for September 12-19, 2026

Jo Van Eyck

Jo Van Eyck

— 8 min read
Share
Agentic Engineering Weekly for September 12-19, 2026

A stealth lab refreshes interest in narrow AI. Agentic coding keeps stressing CI platform infrastructure. The shape of the software factory is converging. Fundamentals matter more than ever. Pacing the frontier is a diversion tactic.

My top 3 picks this week

A new model "Jev" returns structured decisions instead of text and that reopens a design space

TypeSafe AI came out of stealth with Jev, a model that does one thing and does it both fast and cheap. You hand it a set of multiple-choice questions and it hands back typed decisions with calibrated probabilities, in a fraction of the cost and speed a typical LLM would require. As a chatbot it is useless. As a component inside your own code or harnesses it is something we have not had before: intelligence cheap enough and fast enough to wire inside a hot loop.

Dubbed a "System One" model (a metaphorical nod to Kahneman's System I & System II), it's designed to operate as a fast, low-cost decision-making component within larger systems, rather than as a standalone conversational agent.

The coverage has mostly fixated on the 200x speed claim, which I find is the least interesting part. The interesting question is what kinds of workloads we can restate as a handful of classifier tasks. I'm guessing this will be a big deal for AI engineers and harness builders alike. Anything you Jev can do, an LLM can do as well. The staggering difference in speed and cost is what makes the difference.

Optimizing for human preference optimizes for engagement and overpromising. RLHF-built LLM's are extraordinarily good at pleasing the human in the loop, which is the wrong objective when the consumer is a switch statement in code.

Worth reading:

Several benchmarks shift focus from model in isolation to model + harness

For two years the question that's being repeated ad nauseam has been "which model is best". HarnessTax unasks the question. The team evaluated 21 model-harness pairs across seven models and three harnesses (Claude Code, Codex CLI and Pi) on SWE-bench Lite and Terminal-Bench 2.0. The results: your beloved Claude models may not need Claude Code.

There is a shelf-life problem lurking underneath all of this rapid evolution: Uncle Bob's states that most of his harness has already been obviated. Build today's skill/harness/factories but expect to deprecate them at the next model release. That sounds wasteful until you consider the alternative: a harness tuned for last year's failure modes fights a model that no longer has any of them.

Worth reading:

Agent throughput broke review first, CI is next

Anthropic's own CI job volume went up by a staggering 25x in six months. Build and test speeds everywhere are optimized for human speed. Agents do not sleep.

When the constraint on shipping moves from writing the change to verifying it, every queue downstream of the diff becomes the bottleneck. CI, review, staging, rollout, all of them. Management dreams of turning the lights off entirely, while engineers on complex legacy dim them instead. Don't go dark y'all, every factory needs some lights.

Sean Goedecke argues that developer experience measured in seconds is about to matter again. Sub-second feedback used to be pointless because humans needed a couple of seconds to do some actual thinking anyway. In today's world every millisecond is back on the critical path. Zed is betting further out with Delta, which drops the pull request as the review unit entirely. And Steve Yegge ties the economics together: Distrust is what costs money, and trust is the only thing that buys it back. Trust is incredibly hard to win back once it is (so easily) broken.

My fix for all this? I think a lot of companies could benefit from revisiting their test automation strategy. For a typical app or small-to-mid-scaled service, your full CI should run in under 5 minutes. Interested in how to get there? Hexagons are the bestagons, parallellize and virtualize ruthlessly. Or give me a call, I do this stuff for a living!

Worth reading:

The shape of the factory is converging

The Pragmatic Engineer deepdive on how Codex took over engineering inside OpenAI is the most detailed look yet at a frontier lab running its own software factory. The novelty is not what strikes me, though. It is how closely the whole thing rhymes with what Stripe, Zalando and Ramp and many others have published over the past month. Isolate the work, treat it as a first-class citizen, build it, prove it, ship it.

I have been pulling these published factories apart into their moving parts, and the same primitives keep showing up.

The adoption question is where the debate hasn't settled yet. Warp's Zach Lloyd argues first crawl, then walk, finally run. Addy Osmani makes the point that matters most for anyone reading this from inside a real company: greenfield demos are not the hard case, running agents in a codebase older than the team is a different discipline altogether. My own read is that the logical end state goes from mostly human-in-the-loop to mostly human-on-the-loop to mostly cron-events-and-feedback-signals, and that most teams are trying to skip ahead some steps. That won't end well.

Worth reading:

AI is fun. But don't skip fundamentals

MIT's AI and Education report landed in the same week that Matt Pocock, whose grill-me skill is one of the most-used things in the agentic engineering ecosystem, spent an interview explaining why software fundamentals still matter. Skills scale a practitioner or team. They do not create one. The tools that compress the work also compress the apprenticeship and necessary friction that used to produce senior talent.

The field of mathematics is having it's AI crash-out moment right about now. Almost five thousand mathematicians signed a declaration about AI in mathematics. Solving prestigious problems used to be a reliable signal of who could solve prestigious problems. When this signal breaks every credential built on top of it wobbles. Seniority in software had the same structure. "Shipped a working feature" used to tell you something about the person who shipped it. Now anyone with a keyboard and a subscription can ship faster than the world's best trad-coders.

Worth reading:

"Do you still read the code?" stopped being rhetorical

Several old-hats weighed in this week. Martin Fowler published a short piece titled "I don't like LLMs" and was careful to frame the dislike as aesthetic rather than technical, which is more interesting than another capability argument. Ron Jeffries poked at the same thing.

Several experienced engineers (including myself) report that they have stopped writing most of their own code, ship more than they ever have. Some of us like this new style of working, others don't experience the same levels of enjoyment anymore. The small implementation wins that used to create flow are gone. For some people, productivity and satisfaction have come apart. Anyone nudging their engineering teams towards a software factory has to deal with this reality.

Charity Majors supplies the framework: Sometimes AI involvement is additive, sometimes it is a violation, the line moves with the context. Choosing between "use it for everything" or "it is all slop" is a false dichotomy, the productive truth lies somewhere in between.

Worth reading:

More on pacing the frontier

Dario Amodei published "We Must Pace the Frontier". Sam Altman agreed. Zuck tells other frontier labs to stop whining and take responsibility for their product.

Ed Zitron's objection: If you truly believe you are building the dangerous thing, stop building it. We don't need to invent future horror scenario's, we are already boiling the oceans today. Cory Doctorow states that we should separate the technology that works from the investment story wrapped around it. Both of those counterarguments do more work than yet another round of p(doom) estimates in the mainstream media.

Worth reading:

Quick Hits

Curated from 61 sources across articles, podcasts, and videos. Week of September 12-19, 2026.

Read more