TL;DR. In April 2026 an engineer at Anthropic pointed Claude Code at a class of API errors nobody had owned for years. It shipped over 800 fixes and cut that error class by a factor of a thousand, work the engineer estimated at four years by hand (Anthropic, June 2026). The point is not speed. That work was never going to be scheduled at all. It sat below the line where a human's time is worth spending, and that line just moved. This is non-production: the supply-side mirror of the non-consumption argument. What actually collapsed is not labour hours but the cost of holding unfamiliar context, a transaction cost in Ronald Coase's sense, which is why migrations and cleanup get unlocked first. The same force makes over-building the default, so the strategic logic inverts: when building gets cheap, not building gets valuable, and the winner is whoever selects best.
In April 2026 an engineer at Anthropic pointed Claude Code, Anthropic's in-house AI coding agent, at a class of API errors nobody had owned for years. It shipped over 800 fixes and cut that error class by a factor of a thousand. The engineer's estimate for the same work by a human: four years (Anthropic, June 2026).
The point is not speed. It is that this work would never have been scheduled at all. It sat below the line where a human's time is worth spending. That line just moved, and almost no one has priced in what sits underneath it.
We have spent two years here talking about non-consumption: the customers a bloated SaaS stack over-serves and prices out, reachable now with something small enough to actually buy. The Anthropic number points at the mirror image on the supply side. Call it non-production.
When Building Gets Cheap, Not Building Gets Valuable
Non-production is the work that, at normal time and normal cost, was never going to get built. Deferred bug fixes. The migration everyone postpones. Documentation. The exploratory tool that might pay off. Technical debt that is real but never urgent enough to fund. None of it was worthless. It was simply below the hurdle rate, where the value was real but the cost of a human doing it never cleared the bar.
It helps to be precise about what moved. The binding constraint was never raw labour hours. Read the Anthropic engineer's own words: solving other people's bugs is slow "because humans struggle to hold that much unfamiliar context in their head at once." That is a transaction cost, in the sense Ronald Coase meant it. The expensive part was acquiring and holding unfamiliar context, not typing the fix. agentic engineering, software work where AI agents read a codebase, run tests and chain many small fixes on their own, collapses exactly that cost, which tells you precisely which work gets unlocked first: high-context, low-glory, fragmented work. Cleanup, migrations, cross-cutting fixes. The deferred-maintenance pile, in other words. Anthropic engineers now ship roughly eight times the code per quarter they did from 2021 to 2025, and Claude authored more than 80% of merged code by May 2026 (VentureBeat, May 2026).
Here is the part that should make a CEO uncomfortable, because it is the same force pointed the other way. The thing that frees up valuable deferred work also makes over-production the default. When the marginal cost of a feature falls towards zero, the over-engineered product I have been criticising for two years stops being a management failure and becomes the path of least resistance. Non-production is not only the cure for bloat. One level up, it is the cause of it. vibe coding a feature in an afternoon, shipping on instinct with little architectural discipline, does not ask whether the feature should exist.
So the strategic logic inverts. When building gets cheap, not building gets valuable. The cost of bloat does not disappear when production cost collapses, it relocates: cognitive load, maintenance surface, security surface, the review throughput of the humans who still have to sign off. The scarce resource is no longer production capacity. It is the discipline to decide what not to ship, and the judgement to tell a thousand-times improvement apart from a thousand things nobody asked for. The winner is not whoever produces most. It is whoever selects best. In practice that discipline is mundane plumbing: a named owner empowered to say no, a default-no on net-new surface area, and one standing question in every roadmap review, what did we choose not to build this quarter, and why.
The Monday Morning Test
Take your portfolio company's or your team's backlog and split it in two. Column one: deferred work that is genuinely valuable and was only ever blocked by the cost of human context, the migrations and the cleanup. That is now cheap, fund it. Column two: everything you would build merely because you now can. Put a harder gate in front of that column than you had last year, not a softer one. The firms that lose the next cycle will not be the ones that produced too little. They will be the ones that mistook cheap production for a reason to produce.
If this analysis prompts a conversation about where your firm sits in the build-versus-select question, that is what drfloriansteiner.com/angebote is for.