Two posts on the Hacker News front page on October 18, 2026 approached the same worry from opposite ends. One is a tool that forces a developer to understand code an agent wrote. The other is an essay by the chief executive of Linear about what is lost when nobody does.
Sphinx: the push waits until you can explain the change
Sphinx's README opens with its thesis: "You can outsource your thinking, but not your understanding." The mechanism: "When you push to a protected branch, Git waits while Sphinx asks you to predict what your change will do. If you miss something, it explains and gives you a new scenario to try. Once you pass, the same push continues. This works whether you or a coding agent ran git push." The questions do not come from the agent that wrote the code: "The questions come from a separate reviewer model. It reads the committed change through read-only Git tools and does not see the conversation that produced the code." The README's example asks whether a background job can run twice if the worker commits but crashes before acknowledging; the wrong answer gets a one-paragraph explanation of the redelivery window, then a new scenario about a visibility timeout.
It works with Claude Code, Codex, Cursor, OpenCode and Pi or with a plain git push, and the reviewer can run on an existing Claude Code or Codex plan, on OpenRouter, or on any OpenAI-compatible endpoint. The project site adds that it is "Built to run on minimal tokens", that answers never reach the coding agent, that .env files and keys are never opened, and that an escape hatch, sphinx yolo, lets urgent pushes through with every skip logged. It needs Node 22.19 or later and installs with npm install -g sphinx-understand. The repository, under the GitHub user parthh01, was created on October 14, is MIT-licensed, and had 113 stars by October 19.
The Show HN post, by a user named critiquebrowser, reached 242 points and 85 comments. The poster described it as something that "plugs into your existing coding agent workflow" so that "every time you push to prod it'll throw up a questionnaire". The thread split three ways. Supporters saw the real cost: "Learning is the bottleneck for shipping code," wrote one. Sceptics saw the gaming problem: "Nothing prevents a lazy person from copy-pasting the question to the agent," wrote another, and one team lead said his team already bypasses a similar hook. A third group argued about placement, suggesting the check belongs at pull-request time, before "requesting another human's time or attention", rather than at push.
Saarinen: the factories and the creator
Karri Saarinen, co-founder and CEO of Linear, published "Someone behind the software" on October 15; it reached 326 points and 137 comments on October 18. "I think good products carry something of the people who make them," he writes. His concern is not that agents write bad code but that they change how products accrete: "Individuals work with agents to produce changes independently, each operating a small software factory." The result is additive: "A feature can work perfectly well on its own and still make the product worse." And: "Without the shared perspective, soul, or thought, the product becomes an average of all those outputs." He closes: "I want to feel the creator behind the software, not the factories."
The two posts sit in a week of related evidence. Anthropic's October 15 account of reviewing Claude Code, covered here on October 18, found an agent reviewer and an agent author looping for more than fifty rounds and now asks the author to rethink after three. The October 9 Trend Watch covered a 181,711-test compiler port whose author had "never read a line of this code." Sphinx is the first tool in this digest that treats the human's understanding, rather than the code's correctness, as the thing to gate on.
What remains uncertain. Sphinx is five days old and has published no evidence that quizzed developers ship fewer defects or retain more. A reviewer model can be fooled, and so can a developer who pastes the question back into their agent, as commenters noted. We have not verified that the Hacker News poster and the repository owner are the same person. Saarinen's essay is an argument from experience, not data.
Analysis: Comprehension gates are a new category, and the Hacker News objections are also the design brief: make the check cheap, place it where attention already gathers, and log the skips so a team can see how often it is bypassed. Whether a quiz is the right form is an open question, but the problem it names is not. Saarinen's essay describes what a codebase of unread, individually correct changes becomes, and Sphinx is one small attempt to make at least one person read.
Sphinx, an MIT-licensed Git hook created October 14, pauses pushes to protected branches while a separate reviewer model, reading only the committed diff, asks what the change does under failure and keeps asking until the developer answers correctly; its Show HN reached 242 points and 85 comments on October 18, split between those who see learning as the bottleneck and those who note the question can be pasted back to the agent, the same day Linear CEO Karri Saarinen's essay arguing agent-built features average a product into nothing reached 326 points.