Throw away ten times more than you ship

TL;DR: A factory that keeps everything it makes is broken. Once making a thing costs almost nothing, the part worth building is the gate that bins most of the output, and a high discard rate is the sign that the gate works.

The insight

Anyone raised on lean manufacturing knows a scrapped part is money lost and the goal is a line that wastes nothing. That is right only while a part costs real money. It is wrong once a part costs almost nothing, which is where an AI factory sits. Cheap output moves the constraint from making to choosing, and a gate whose job is to say no does the choosing. We wrote about the making side in Ten times faster at eight per cent of the job; this is about what happens after.

The danger is what a founder does with the bin. If waste means failure, a full bin looks like a defect, and the quiet fix is to loosen the gate until more gets through. That is backwards. The gate is the one check standing between cheap output and the product, so weakening it turns a healthy discard rate into a stream of unreviewed change. Build the gate first, make it bin by default, and read a rising reject rate as the gate doing its work.

In practice

Mik Kersten’s book Output to Outcome: An Operating Model for the Age of AI describes Cognition, the company behind the Devin coding agent, running up to ten rewrites of its own product at once. The founders see them at a Sunday review, keep the few that earn it and drop the rest. Kersten, citing swyx at Cognition, says the company bins ten times more code than it ships. We took his point, that making too much lets you learn faster once output is cheap, and applied it to prose and research findings as well as code.

Our own evidence is an internal research note, written in July and not yet published, on putting a second vendor’s coding agent on a second machine. It reached its final draft with three of its confident findings thrown out and replaced by measurement. A DNS allowlist that looked like a security control was Peter’s home router failing to resolve two domains. A sandbox we believed a flag had disabled had never been switched on. And OpenAI’s Codex agent, which we assumed could not intercept its own actions, turned out on a reading of OpenAI’s own documentation to have eleven hooks that can. We trust the note because it records what it discarded.

What we’ll try next

The loop that reads each day’s logs and proposes changes to the factory’s own rules reports how many proposals it made and how many landed. Next it will report its reject rate as a number, in the same place, every week. A week in which that loop declines nothing is an alarm that the gate has gone soft, not a triumph that every idea was good.

One honest number

Ten to one. Kersten reports that Cognition at times throws away ten times more code than it puts into production, and we now take that ratio as the target shape for our own gates. Ten is not magic; the point is that a gate which keeps most of what arrives is a gate that is not choosing. Assume an unverified claim is wrong until you have measured it, and treat a gate that is rarely busy as open rather than strict.