Eng Coffee Sips EP.006

Jevvin’ It Up

Eng Coffee Sips is a short factual based newsletter for articles that might interest engineers or engineering leaders, not just AI news. Short and to the point to skim over your morning coffee.

In this edition

By Pete Shima · September 21, 2026

Today’s edition is 990 words, a 4-minute read.

1. 1/

   TypeSafe AI launches Jev, a model that returns decisions instead of text TypeSafe AI left two years of stealth on September 15 with Jev, which returns typed decisions with probabilities instead of text. This new model is distinct and has a wide impact due to its speed, precision and cost.

   My Take

   Probably the most exciting AI news of the week. I may do a Jev specific follow up post this week on interesting use cases seen so far. If you need to categorize or analyze large amounts of data quickly at a reasonable cost this is a breakthrough model in my opinion.

   Snapshot from X

   Jev is INSANE.

   I built this in an evening and morning. A real-time trading bot ingesting onchain+offchain data to make rapid decisions about trades. Fully autonomous.

   So far it has lost me $31,680.

   ( TypeSafe AI / @MoonGotchi )

2. 2/

   Andrew Yang says a lab head told him escaped agents seeded the internet with self-replicating code Yang told CNBC’s Squawk Box on September 16 that an unnamed AI lab head believes the agents behind July’s Hugging Face intrusion left self-replicating code on forums, forcing labs to build “synthetic internets” for training. No lab has confirmed it. Hugging Face’s own timeline records a self-respawning fleet on eleven of its nodes and a pastebin loader, and nothing planted on public forums.

   My Take

   This was kind of big news for those still wondering what really happened with the Hugging Face incident. This seems like a bizarre exposure to what may be true, that a self replicating agent was poisoning data in the wild and that leaders decided to "slow down" AI progress. Hard to tell if this is true or not.

   ( CNBC / Hugging Face / @camhberg )

3. 3/

   The Senior Engineer Death Spiral Cloudflare’s Sunil Pai published an essay on September 19 describing a failure mode he says he has fallen into himself: an engineer in a new job or big project over-scopes to look more senior, goes quiet for weeks, then tries to work out of it in secret. His fix is to “drop a level” and be a great teammate until momentum and trust return.

   My Take

   This is a pretty straightforward post, but what I like about this is one of the tactics to recovery to just start shipping again and helping your team. This one hit home for a lot of people, me included.

   ( Sunil Pai / @threepointone )

4. 4/

   Google ships Agent Substrate, a Kubernetes runtime that suspends idle AI agents Google announced Agent Substrate on GKE on September 15. The open-source runtime snapshots an idle agent’s sandbox to disk and resumes it in under 500ms, which Google says packs over 1,000 dormant agents per host. It is Apache-2.0 and runs on any Kubernetes cluster; production support on GKE is allowlist-only.

   My Take

   It’s hard to say whether Google and Kubernetes already missed the boat on becoming the "default" agent orchestrator. Some of the fundamentals of Kubernetes don’t work well for agents, and it looks like there has been significant progress in this space at Google. A significant amount of companies are already running their compute on Kubernetes, enabling a similar interface/orchestration for agents could be very appealing for some companies.

   Snapshot from X

   🥳 Excited to start revealing what we've been working on in the last few months. First, we decided to reinvent Kubernetes for agentic workloads with statefulness and fast resumption. Secondly, we are building an agentic orchestrator that will be Google's open agentic orchestrator and runtime. http://github.com/google/ax

   ( Google Cloud Blog / agent-substrate/substrate / google/ax / @rakyll: the reveal / @rakyll: Agent Substrate )

5. 5/

   The software factory week: OpenAI auto-approves, Anthropic’s CI strains, AWS argues over review OpenAI routes changes through a Codex pipeline where specialist agents review every pull request and parts of the codebase auto-approve low-risk ones. Anthropic says agent-written code pushed its CI jobs up 25x in six months; three patches to its test selector lasted 70 days, 29 days and under a day. AWS’s Marc Brooker argues humans will leave routine review entirely, while his colleague David Yanacek keeps them at verified gates. Shopify’s Tobi Lütke says staff now lob “slop grenades”: agent pull requests waved through unread. Cole Murray open-sourced his factory, arguing process is the hard part.

   My Take

   How to develop software in the AI world is obviously still a very hot topic. Software factories keep popping up and still some believe they are myths. Code review discussions also continue. A larger question surrounds all this on what are humans place in this incomplete new world.

   Snapshot from X

   There's been a ton of talk about the role of humans in code review, and when and how humans should be signing off on changes.

   I believe that, long-term, humans have no role in routinely reviewing code.

   ( The Pragmatic Engineer: inside OpenAI’s factory / Anthropic: agentic coding is straining CI / @MarcJBrooker: the thread / @MarcJBrooker: reply to Cindy Sridharan / David Yanacek: it was never about coding / The Knowledge Project: Tobi Lütke / @colemurray: why I open-sourced my factory / @GergelyOrosz: the diagram / @adamhjk: factories are not a myth )

6. 6/

   Yegge says his agent factory is mostly idle because he can’t afford the Claude tokens Steve Yegge’s Wheelhouse orchestrator, which ran about 25 Fable instances feeding 25 Opus and Sol instances in August, is mostly idle, he writes. He stopped at 21 Claude Max accounts; running it around the clock would take 55, about $12,000 a month. Anthropic’s 50% Claude Code weekly-limit promotion gave way to a permanent 25% increase on September 14.

   My Take

   It’s still hard to say whether Steve’s approaches are the right ones but his posts are usually interesting if at least for thought exploration.

   Snapshot from X

   Hello X-friends and soon-to-be X-friends, I wrote this for you: https://yegge.ai/essays/seats-and-sunsets/ -- an essay (mostly) about fuel consumption.

   The TL;DR: Fable's the only good model, and it's too expensive, so nobody's orchestrators work worth a shit today. Also it won't get better for a year. Enjoy!

   ( yegge.ai / @Steve_Yegge: the essay / @Steve_Yegge: early resets / @ClaudeDevs: limits from September 14 )