Five OSS authors on a 48K Spectrum demo, a one-command toolchain, and the end of line-by-line review (2026)

Four of this week's five statements are maintainers deciding that work they once did by hand is worth handing over, and naming what should absorb it. David Heinemeier Hansson says reading every line of agent-written code is a practice he expects to end, and that adversarial review plus automated tests have to replace it 1. Evan You shipped a toolchain whose whole point is that a team stops assembling one 2. Armin Ronacher argues the escape hatch Rust programmers are taught to keep shut has become affordable 3. Salvatore Sanfilippo ran a 1991 game's virtual machine on a 48-kilobyte home computer, directing a model rather than a team 4. Simon Willison, who works with the same tools daily, reports the cost from the other side: he finds them making software engineering harder, and raising the premium on discipline 5.

Five choices in one view

Author	Fresh signal	Stance in the source	Design question	Question for a tech lead
David Heinemeier Hansson, creator of Ruby on Rails and chief technology officer of 37signals	September 28 post on the review work agents add 1	Reading every line of agent-written code is a practice he expects to end, and adversarial agent review, automated tests, and spot checks are what remain.	Which parts of review exist to catch what a test could catch, and which hold judgment a test cannot encode?	If your team stopped reading agent diffs line by line tomorrow, which check would fail first, and how long before you found out?
Evan You, creator of Vue.js and Vite, founder of VoidZero	September 28 release of Vite+ 1.0 and the Cloudflare post behind it 2 6	Maintaining a tool stack is overhead a team should be able to buy its way out of, and the toolchain's own latency now sits inside the agent's feedback loop.	When one toolchain serves both people and agents, whose latency budget does it sit in?	How many config files, upgrade guides, and internal scripts does a new repository inherit in your organization, and who keeps them alive?
Simon Willison, co-creator of Django and creator of Datasette	September 24 note on coding agents 5	The tools raise the ceiling and the floor at once: more is possible, and unlocking it costs extraordinary discipline and knowledge.	Where does a team's understanding of a system live once the diff stops being where it accumulates?	Which of your engineers can describe what the last agent-authored change did, without opening the diff?
Salvatore Sanfilippo (antirez), creator of Redis	September 28 release of an Another World port for the 48K ZX Spectrum 4 7	One person directing a model can now finish work that used to need a team, and the skill that matters has changed rather than disappeared.	Which features in a roadmap wait on headcount that will never arrive?	What could one engineer plus an agent finish this quarter that your roadmap has parked for years?
Armin Ronacher, creator of Flask, developer of the Pi coding agent	September 25 post on Rust and unsafe code 3	A constraint that existed because verification was expensive is worth re-pricing once agents can do the verifying.	Which of a codebase's conventions was a convention only because checking used to be costly?	Which rule in your style guide has nothing left to stand on except that it was true before agents?

David Heinemeier Hansson: the review that stopped paying

David Heinemeier Hansson, the creator of Ruby on Rails and the chief technology officer of 37signals, wrote on September 28, 2026 that reading agent output line by line cannot hold as a practice 1:

"It takes a while to build the confidence, but there's no future where you're manually reviewing every line of agent code. Not much acceleration in that. You need adversarial agent reviews, you need automated testing, and maybe you spot check. That's it. From prompt to production!" 1

Loading content card…

His replacement has three parts and no fourth: a second agent whose job is to attack the first one's work, an automated test suite as the gate, and human spot checks where suspicion runs highest. The claim rests on a judgment he posted the day before, that the current models had crossed a line for him — "The hype is true. Anthropic did magic with Opus 5.5. Damn it's good." 8 He put the consequence for working engineers in the same terms that afternoon: "The only programmers who are really in trouble with the competition from artificial intelligence are the ones who insist the world today isn't all that different from the one we lived in last year." 9

Once the diff stops being the review, the process carries the guarantee, and a process can be measured in ways a reader cannot. That is what makes his three parts a system rather than a slogan: the adversarial agent and the test suite produce evidence, and the spot check spends the one resource the other two cannot replace.

Evan You: stop maintaining the toolchain

Evan You, the creator of Vue.js and Vite and the founder of VoidZero, released Vite+ 1.0 on September 28, 2026, with three colleagues on the announcement 2. Vite+ places one command, vp, in front of nine jobs: scaffolding a project, installing dependencies, running the dev server, formatting and linting and type-checking in a single pass, running unit and browser tests, building an app, building a library, running monorepo tasks with caching, and managing the Node.js version 2. One vite.config.ts at the project root configures all of them. The release is MIT licensed, and the announcement is careful about what it is: "neither a framework, nor a package manager and definitely not a replacement for Vite" 2.

The cost it targets is the one that grows with the number of repositories:

"Teams with more than a handful of repositories usually end up building internal scripts to paper over this, which then need their own maintainers. But instead of maintaining a tool stack, you often just want to ship features." 2

A single vite-plus dependency replaces the cluster most teams install — vite, vitest, eslint, prettier, tsup, turbo — together with their plugins and configs, so an upgrade becomes one version bump tested as a unit 2.

The second argument for the release concerns who is waiting on the toolchain. In the Cloudflare post published the same day under his byline, You writes:

"when we optimize for developer experience, we no longer just optimize for humans — we optimize for agents. And as inference gets faster, the speed of type-checking, linting, or building the code becomes the bottleneck again. The longer those processes take, the longer an agent has to wait before making progress and completing its goals." 6

The same post counts the four months since VoidZero joined Cloudflare — more than 80 releases and over 1,200 issues closed — and records a $1 million commitment to a Vite ecosystem fund and a second $1 million to open source 6.

Simon Willison: the discipline the tools demand

Simon Willison, co-creator of Django and creator of Datasette, published a two-sentence note on September 24, 2026 5:

"The more time I spend working with coding agents, the more convinced I am that they make software engineering even harder. We can do amazing things with them, but unlocking their full potential requires extraordinary discipline and knowledge." 5

He posted the same line on X the next day as an answer to Gergely Orosz, who had argued that non-developers will not be shipping production code for a long time 10 11. Willison accepts that conclusion about who ships and moves the difficulty somewhere else: the same toolkit that produces a working change also produces a change the team has built no model of, and the discipline he asks for is what keeps the first from becoming the second.

He is working in these tools himself while saying it. That same week he pointed an agent at the problem of reply bots, and shipped a Bluesky profile checker that looks for answers posted seconds after someone else's post, accounts that only ever reply, and question marks in bot replies 12. The note and the tool describe the same arrangement: the capability is real, and using it well is a skill the team has to hold.

Salvatore Sanfilippo: one person on a 48-kilobyte machine

Salvatore Sanfilippo, who writes as antirez and created Redis, published a working port of Another World for the ZX Spectrum 48K on September 28, 2026 7. The repository states the division of labour in one line: "A Z80 port of the actual Another World virtual machine, by antirez, operating GPT6 Astra." 4

Another World, the 1991 game by Éric Chahi, drew its polygon graphics by running its own virtual machine, so a port means reimplementing that machine and feeding it the game's bytecode 4. The target is an 8-bit home computer with 48 kilobytes of memory. The README states that both the virtual machine and the intro data fit inside those 48 kilobytes, with no streaming from tape, and it lists the work that bought the space: one-byte references for palette and text commands, coordinates packed two to a byte, the hardware screen doubling as the front buffer, strings and palettes hidden in the invisible pixels of the screen margin 4. Copyright in the port's code is assigned to Chahi, and the artwork and script remain his 4.

His reading of the week's argument is that the skill changed rather than disappeared 13:

"Good programmers state they use models like GPT 6 Astra without obtaining good results: it makes me feel I live in a parallel universe. But the explanation is not that hard: good programming in the past and now requires a different skill set, even if there is some overlap." 13

Four days before the release he gave younger programmers the same instruction as an opening move: "Create open source software that does not exist currently since in the past it required teams of people and there were not the economic incentives but now with AI it is possible for individuals to do it." 14

Armin Ronacher: the constraint the agents made affordable

Armin Ronacher, the creator of Flask and a developer of the Pi coding agent, posted on September 25, 2026 3:

"Now that agents are this good, Rust developers should stop being afraid of unsafe." 3

Loading content card…

unsafe is the keyword that unlocks what Rust's compiler will not verify on its own: raw pointer work, calls into C, memory managed by hand. The convention around it has been to confine each block and audit it by hand, because every one of those blocks moves a guarantee off the compiler and onto the person reading the code. Ronacher's claim re-prices that audit. An agent that can re-read, rewrite, and test the block lowers what the human review costs, and a rule inherited from a time when verification was expensive deserves to be checked against today's price.

The post is one line and its 53 replies show how much the trade-off is still argued over, which is the useful part for anyone holding the same rule in their own codebase. The claim is narrow — Rust's unsafe keyword — and the reasoning behind it generalizes to any convention whose only justification was that checking used to cost a lot.

Questions for the next design review

Review substitution: if the team stopped reading agent diffs line by line tomorrow, which check would fail first, and how soon would you know?
Toolchain ownership: how many config files, upgrade guides, and internal scripts does a new repository inherit, and who is on the hook for keeping them current?
Understanding: where does the team's model of a system live now that the diff is no longer where it accumulates?
Individual scope: what could one engineer plus an agent finish this quarter that the roadmap has parked for years?
Inherited rules: which rule in your style guide has nothing left to stand on except that it was true before agents could check the code?