{"slug": "generation-went-to-zero-somebody-still-has-to-read-it", "title": "Generation Went to Zero. Somebody Still Has to Read It.", "summary": "Linear's CTO detailed how agentic coding drove the company's test suite to nearly quadruple, with roughly 2,000 new tests added weekly, forcing months of pipeline performance work to keep pull request wait times from growing. Shopify's CEO flagged the resulting \"slop grenades\" — AI-generated work passed to colleagues without evaluation — as the real cost of faster generation, while a tool called Foremerge targets silent merge failures between agents working in isolated worktrees.", "body_md": "The cost of producing software didn't vanish this year. It moved to places with worse budget lines: your runner minutes, your merge queue, and the finite attention of whoever has to read the thing.\n\n## 1. Agents Made Shipping Fast. Then Everyone Started Waiting in the Same Line.\n\nLinear's CTO opened the year by assigning an engineer a ticket titled \"CI costs are high\" — and, while they were at it, asked for CI to be faster too. The resulting [rework of their pipeline](https://linear.app/now/ci-bottleneck-reworked) is the clearest accounting I have seen of what agentic coding actually costs. Agents made producing changes exponentially faster; validating them did not speed up at all. Their test suite nearly quadrupled over the year — they add roughly 2,000 tests a week now — and the only reason pull request wait time fell from over six minutes to just over five is that a team spent months fighting that growth to a draw.\n\nWhat they did is striking for how little of it is about AI. The native TypeScript compiler cut the median typecheck by 73%. Capping fetch depth on a change-detection job that gates eight test shards took it from 94 seconds to 20. Consolidating seven tiny checks — each booting a runner and installing dependencies to do seconds of real work — into two jobs saved about 87,000 runner-minutes a month, 11.8% of CI usage. Unglamorous performance engineering. What changed is that it became urgent.\n\nThe detail I keep returning to is at the end. Their biggest win was letting safe test files share a module registry instead of isolating every one — and because agents now write most of their tests, they updated their agent skills so generated tests honor that opt-in. The pipeline's constraints had to become context the model reads.\n\n**Why it matters:**\n\n- **For ICs:** profile the gates, not just the tests. The 94-second job nobody looks at is holding up eight shards.\n- **For leaders:** CI spend is the lagging indicator of agent adoption. If you approved the assistants and not the platform work, you approved half a plan.\n- **For founders:** \"we ship faster with agents\" is only true through merge. Measure time-to-merge, or you are measuring the part that got easy.\n\n## 2. Half of Shopify's Pull Requests Are Machine-Written. The CEO's Complaint Is About Email.\n\nShopify has gone about as far into this as anyone — its internal agent reportedly handles as much as half the company's production pull requests. So it is worth noting what its CEO flagged [as the actual problem](https://www.businessinsider.com/shopify-ceo-ai-slop-grenades-can-make-work-harder-2026-9): AI lets people produce more material, faster, without taking responsibility for whether it is useful or correct. Internally they call the results slop grenades — work you toss at a colleague without having evaluated it. His example is an AI-written email long enough that the recipient runs a model to summarize it. \"Why did we invent decompression and recompression? This is terrible.\"\n\nThe mechanism under the joke is a cost transfer. Effort used to be a signal: a long document meant someone thought it was worth a long document. When producing costs nothing and evaluating costs what it always did, the rational individual move is to send more, and the result is queues full of work nobody checked. No performance review catches this, because the cost lands on someone else's calendar.\n\nAgents do it to each other too, in a form Git cannot see. A coordination tool called [Foremerge](https://github.com/naw103/foremerge) exists for exactly one failure: two agents in isolated worktrees, no overlapping lines, both changes individually correct, and the merge quietly destroys the work. One moves every caller onto a new StripePaymentService while the other adds PayPal support to the old PaymentService — nothing conflicts textually, so both land, and the PayPal work is stranded on a class nothing calls. Its answer is to make agents declare what they intend to touch before they touch it. At version 0.5 with no benchmarks, treat it as a sketch; the diagnosis is what lasts. Git compares text, not intent, and we just started running many authors in parallel who share none.\n\n**Why it matters:**\n\n- **For ICs:** if you would not sign it, do not send it. Reading is the scarce resource on your team now.\n- **For leaders:** name the behavior before it becomes culture. An org where forwarding unreviewed AI output is acceptable will run slower than one with no AI at all.\n- **For founders:** parallel-agent coordination is an open category. Nobody has a good answer for intent conflicts, and everyone running more than two agents will need one.\n\n## 3. Your Reviewers Are Already Bailing Out, and They Have the Numbers\n\nAn engineer's essay this week put the human side plainly: [I don't want to read what you didn't write](https://blog.colinbreck.com/i-dont-want-to-read-what-you-didnt-write/). The pattern will be familiar — someone builds a thing with AI, then has AI retrospectively summarize it into a design document. That document is no longer a proposal meant to build consensus; it is an exhaustive machine summary with no perspective. Same for pull request descriptions written by machines for machines: rich in what changed, silent on why and on where the author wants input.\n\nHis explanation of why it lands badly is the sharpest part. When you prompted the model you already hold the context, so you can skim the output and discard what is wrong. The reader cannot skim; they must read exhaustively, hoping to establish context. The survey numbers he cites are brutal: if readers believe an article is AI-assisted, 78% stop reading and 71% avoid the author in future, while 98% prefer the author's own writing with all its flaws. Using a model to write, as he frames it, voids the contract between writer and reader — we shouldn't labor over a sentence the writer didn't labor to create.\n\nWhat makes the essay credible is that he used AI heavily on an academic paper and it wrote none of it. It verified his paragraphs against source code and production metrics while he kept writing, filled in citations, and caught a notation error four expert reviewers had missed. Asking it to write the paragraph from that same context was never once valuable. That is the day's theme in one workflow: pointed at verification, the model is a superpower; pointed at generation aimed at a human, it manufactures work downstream.\n\n**Why it matters:**\n\n- **For ICs:** write the pull request description yourself, even badly. Three sentences of real intent buy more review quality than forty lines of generated changelog.\n- **For leaders:** reviewers quietly ejecting from AI-shaped documents are not being precious. They are protecting the scarcest input you have left.\n- **For founders:** verification tooling is the underserved half of this market. Everyone sold generation; almost nobody sold the check.\n- The pattern across all three: cheap production plus unchanged verification is not a productivity gain. It is a queue forming where you are not measuring.\n\n## The Verdict: Real or Hype?\n\n**CI as the new AI bottleneck → Real.** Validation is the one thing agents did not accelerate, and it is now everyone's critical path. **AI-written prose aimed at human readers → Hype.** Three-quarters of your audience ejects on detection, and the decompression tax is paid by people who never chose it. **Intent coordination between parallel agents → Real but early.** The failure mode is real and Git structurally cannot see it; the tooling is version 0.5.", "url": "https://wpnews.pro/news/generation-went-to-zero-somebody-still-has-to-read-it", "canonical_source": "https://fromtheterminal.substack.com/p/generation-went-to-zero-somebody", "published_at": "2026-09-22 22:10:23+00:00", "updated_at": "2026-09-22 23:24:20.664747+00:00", "lang": "en", "topics": ["ai-agents", "ai-tools", "developer-tools", "mlops"], "entities": ["Linear", "Shopify", "Foremerge", "StripePaymentService", "PaymentService", "PayPal", "TypeScript"], "alternates": {"html": "https://wpnews.pro/news/generation-went-to-zero-somebody-still-has-to-read-it", "markdown": "https://wpnews.pro/news/generation-went-to-zero-somebody-still-has-to-read-it.md", "text": "https://wpnews.pro/news/generation-went-to-zero-somebody-still-has-to-read-it.txt", "jsonld": "https://wpnews.pro/news/generation-went-to-zero-somebody-still-has-to-read-it.jsonld"}}