Eng Coffee Sips EP.009

No Code? No Problem

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 · October 12, 2026

Today’s edition is 1806 words, an 8-minute read.

Does anyone write code anymore? A week of engineers arguing about what the job is now Lauren Tan argued across four posts that every codebase has always held slop, that Good Code means correct, bug free and cheap to run, and that readability was a cost paid for human maintainers. DHH replied “Agree.” Gergely Orosz said engineers who refuse AI “won’t be hired at 99.9% of startups+Big Tech+ambitious companies”. Sunil Pai said engineers never liked hard problems, only puzzles, and a “cheat code machine” now does those.

My Take

AI "can’t do code reviews", or AI "can’t write code as good as humans" and other arguments are still standing ground but I wonder how much of this will be backed by how software development "used" to be rather than what it will actually become. Many companies are driven by cost, and this is a cheaper way to achieve the goal. The code they had written by engineers wasn’t the best. Why? Because the requirements changed over time and the system was no longer designed for what it was currently doing. You couldn’t rewrite everything all the time. But can you now? Maybe. But what about in 6-12 months. If the code is ultimately worse, but you can rewrite it every 3-6 months, or build a more generic flexible system, is it a worse outcome? Lots of nuance and we’re not going to solve it with this post. There’s a new age of software development coming and not everyone is going to come along for the ride.

( @poteto: codebases have always contained slop / @poteto: clean code was for humans / @dhh: Agree / Geoffrey Huntley: replaced / @GergelyOrosz / Sunil Pai: do you actually like hard problems? / @threepointone / @adamhjk: most AI adoption is awful / @eniwhere_ / SonarSource: does code cleanliness affect coding agents? )
Reverse engineering, agent-assisted: PS2 discs on a PS5, seven Adobe apps in Rust, and a decompiler for your coding agent PS5SX2 boots original PS2 discs on a jailbroken PS5, and every one of its last 100 commits carries a Claude co-author trailer. PS5 Proton runs Windows games on the console’s own OS. ArtCraft lists pure-Rust reimplementations of Photoshop, Illustrator, Premiere and the rest, “vibe coded” in two days; PhotoCraft passed 50,000 stars while its own roadmap puts parity “well below 50%”. And REA, an MCP server that hands a coding agent Ghidra or IDA, went from 10,000 stars to 154,000 in a week.

My Take

Bespoke and personalized software continue to pop up. The PlayStation case is interesting because players felt like they were lied to. That it was too difficult to run PS2 games on PS5, then someone just vibe codes it. Adobe and Office suite now being rewritten in Rust. Apparently these are still bad and buggy, but these were made in a short time window. Are these AI slop? Maybe but the potential still seems there to me. Someone also rewrote their LG firmware for their TV. Software is becoming significantly more flexible. How far off are we from personalized software for 80% of the software you use? What if software building blocks were there and you could have your own interface? I saw another post I can’t find saying we don’t want to use your SaaS agent, I want to bring my own Agent. I haven’t seen too many "BYOA" use cases yet but as we see with our AI assistant tech, maybe this actually is the reality.

( GitHub: Swordpdf/PS5SX2 / @sword_pdf: launch video / Kotaku / GitHub: mihawk-99/PS5_Vulkan / @Mihawk9_9: PS5 Proton / ArtCraft: Crafting Apps / GitHub: storytold/photocraft / PhotoCraft roadmap: honest parity assessment / @echelon: vibe coded seven Adobe apps / Hacker News: ArtCraft Apps / GitHub: morluto/rea / rea.tools / Hacker News: REA / @MatthewBerman )
Epic says no more hand-written code, and the industry hears “it is over for non-AI workflows” Tim Sweeney says a “huge pivot” is underway at Epic, retooling developers for “maximum AI acceleration”; the aim is not “the same stuff with fewer people” but more output. Minutes later Bloomberg reported Epic had delayed its Disney shooter and told the team to go “AI first”: “no more hand-written code”. Epic’s COO had posted a game built with Opus 5.5 driving Unreal through its new MCP plugin days earlier. Epic laid off over 1,000 in March. Genvid CEO Jacob Navok, quoting Sweeney: “If you’re not doing the same you will be left behind. It is over for non-AI workflows.”

My Take

Reported that Epic was going AI first and Tim comes out with the full take. I don’t think he’s wrong, but I think if you aren’t reading Twitter how do you find this? The haters are coming out, but so are the positive takes. This is related to the first post. The AI revolution is inevitable and it’s how people adapt to it that will make the difference.

( @TimSweeneyEpic / @JNavok: it is over for non-AI workflows / Bloomberg / Video Games Chronicle / @EpicVogel: a game with no hand-written code / Epic Developer Community: Unreal MCP in Unreal Editor / BNN Bloomberg: the March layoffs )
OpenAI puts Codex on a 28-day clock: ship something every day or reset everyone’s usage OpenAI pledged that for 28 days Codex would ship one improvement for most users every day, or give every paid account a full usage reset. One week in: a ~50% default speed bump, Auto-Review made free, a Mac meeting note-taker, the Decisions API in public beta, and GPT-6.1 Sol’s default context raised from 272k to 500k. Three resets went out anyway, one after a community vote: “it seems that the game is rigged in reset’s favor.”

My Take

This has been pretty hilarious so far. While there have been some good updates, the improvements for most users every day hasn’t materialized and OpenAI has set them up for a long series of resets. But, was that the plan all along? Just gain some attention with the challenge? Maybe, but to see the contradictory statements on context windows is surely puzzling.

( @thsottiaux: the pledge / @thsottiaux: the vote reset / @thsottiaux: day 7, 500k context / @ChatGPT: the Meetings plugin / OpenAI Developer Community: day-by-day index / The New Stack / Codex Resets: the tracker )
Anthropic ships Haiku 5.5 at a tenth of Haiku 4.5’s price, and Claude moves into dashboards, animation and Google Docs Haiku 5.5 is priced 90% below Haiku 4.5 for prompts up to 100,000 tokens and 50% below above that, with a 1M context window and the effort setting. Sonnet 5.5 cache reads were halved the same day. Claude Dashboards queries warehouses from a plain-language question and shows the query behind every number; Claude Motion animates by writing code rather than calling a video model; a Google Workspace add-on edits Docs, Sheets and Slides in place. All three are in beta.

My Take

I still see the posts on ’I dropped Anthropic and Codex is the GOAT’ posts, but in my lived reality Claude still is the champ. Codex does alright but Opus 5.5 is an incredible balance of cost versus effectiveness. Codex still stumbled on very basic and repetitive items. It’s not just the model and harness, Claude Design is still really good and motion and video have significantly improved. The Google Workspace additions for me are also very welcome but I haven’t used them yet. The Claude ecosystem doesn’t seem as complete as OpenAI but I would really struggle to move over to OpenAI.

( Anthropic: Claude Haiku 5.5 / Claude Platform docs: Haiku 5.5 / Artificial Analysis / Claude: Dashboards and Motion / @claudeai: Haiku 5.5 / @claudeai: Dashboards and Motion / Claude: Google Docs, Sheets and Slides / @claudeai: Google Workspace )
Microsoft’s problem isn’t just their model or tech. It’s about trust. Jacob Andreou, who took over Copilot in March reporting to Nadella, says the team is “working relentlessly to improve Copilot, to earn it’s right to exist for each customer”, with “a ways to go”. The same week Microsoft made two bets under it: Maia becomes the primary inference chip for its own products from 2027, claiming 40% more tokens per megawatt than leading GPUs, and MAI-Code-1.1 Flash, its 137B coding model, now runs locally on Windows with no inference charge, if you have more than 120 GB of RAM.

My Take

Jacob seems delusional. Microsoft shouldn’t be working toward the right to exist. They should be working on building products people love. The problem isn’t just with Copilot for me. It’s when I update Windows I’m back at an install screen asking me to purchase office 365 or that I’ve got ads in my search bar. Or maybe its how much they have lost my trust on Xbox and GitHub. The products keep getting worse, not better. Shoving AI into people’s lives is only going to make it worse. I dunno are the folks that were successful at building software primarily used by teenagers going to save Microsoft? I just don’t buy it.

( @jacobandreou: earn its right to exist / Microsoft: Copilot leadership update / Microsoft: Scott Guthrie on Maia / Windows Experience Blog: hybrid intelligence / @satyanadella: Maia )
The Jev clones got benchmarked: Microsoft’s Decision-1 ties Jev, OpenAI’s Decisions API trails Microsoft shipped Decision-1, post-trained from Qwen3.5-9B, at Jev’s list price to the cent. OpenAI’s Decisions API went to public beta. Fabio Akita then benchmarked 23 decision models: Jev and Decision-1 tied at 97.6% accuracy, Decision-1 at a third of the per-decision cost, OpenAI at 90.5%. TypeSafe, meanwhile, closed $870M at $7.5B.

My Take

Microsoft is shipping though, at least in some areas. The Jev wars continue.

( Microsoft: introducing Microsoft-Decision-1 / Fabio Akita: which Jev-like decision model is best? / OpenAI developer docs: Decisions / TypeSafe: Series AI / Cloudflare Blog: Clef, faster, cheaper, multimodal / @satyanadella: Decision-1 / Hacker News: Decision-1 )
Uber publishes the guardrail in front of its customer agents Uber released Persona Guardrail, the runtime defence in front of its customer-facing agents: a Qwen3-30B classifier checks every turn against a per-agent allowlist and blocklist of intents. Against a generic LLM judge it lifts out-of-domain detection from 57% to 94% at about 20 ms per call. The benchmark is on GitHub, as is ADR, Uber’s system for watching coding agents on employee machines.

( arXiv: Persona Guardrail / GitHub: uber-research/page-benchmark / @UberEng / GitHub: uber/ADR )
Oxide raises $445M after turning a profit Oxide Computer, which sells rack-scale on-prem cloud systems on AMD EPYC, closed a $445M Series D led by Eclipse, with AMD joining as an investor. It says it paid income tax this spring on ordinary operations, scaled manufacturing 20x in twelve months, and still has demand exceeding supply. Cantrill: “we certainly did not anticipate the explosion in compute demand from agentic AI.”

My Take

It is nice to see Oxide make it. Maybe taking funding isn’t "making it" but being profitable seems like it is. Enterprise hardware sucks.

( Oxide: our $445M Series D / PR Newswire / Axios / Hacker News / @bcantrill )
Meta and Sierra publish a draft protocol for agents that sign in as you Sierra and Meta are drafting Personal Agent Protocol with Genesys, Shopify, Stripe and Walmart. A company publishes a well-known poppy.json; an agent starts as an identified guest, then upgrades through OAuth to read or write scopes. Draft 0.1 is up under Apache 2.0. Amazon, OpenAI and Anthropic are not on it.

My Take

So while this isn’t that big of a deal, it highlights a point I have been thinking about recently. We’re in the age of personalized software. What are the consequences of that? Well one of those is that standards are going to be a lot harder to maintain over stacks. These sort of mutual agreements across companies seems positive.

( Personal Agent Protocol: draft 0.1 spec / Sierra: Bret Taylor and Clay Bavor / @SierraPlatform / Constellation Research )
Shipper: the two-pizza team is too big now; try two slices Dan Shipper argues Bezos’s two-pizza team is now too large for building software. His replacement is the two-slice team: one or two people, because anything more adds coordination overhead. Every runs four products with one person each, 20 full-time staff, and says 99% of its code is written by agents.

My Take

I wrote an email to someone recently that describes this type of team but I wasn’t clever enough to think of the two slice term. Nice to see this concept in the wild.

( Every: The Two-slice Team / Lenny’s Podcast: the talk / @lennysan )
Sabi raises $50M to put a brain-computer interface inside a baseball cap Sabi raised a $50M seed led by Khosla Ventures at a reported $600M valuation for a baseball cap lined with up to 100,000 non-contact EEG sensors that read through hair. A “Brain Foundation Model” trained on 100,000-plus hours of neural recordings turns imagined speech into text. A prototype previews at CES.

My Take

What in the fuck is this.

( Sabi press release / Forbes / @rahulchhabra07 )