The AI verification burden is now the cost that decides whether agents pay off. A small open source tool that quizzes developers before they can ship code built by an agent reached 252 points on Hacker News in two days, because it names a problem measured data already shows: output rose, finished work did not, and checking filled the gap.
A quiz at the point of shipping
On 18 October 2026, a Show HN post titled "Sphinx – A hook that quizzes you on your AI's code before shipping" introduced a Git hook that stops a push to a protected branch and asks one to three questions about the change, scaled to its size. The questions cover behaviour, edge cases and what happens when something fails. A wrong answer blocks the push and explains the gap. The project description, as reported in coverage of the launch, is plain: "You can outsource your thinking, but not your understanding."
The Sphinx thread argues that comprehension was once a byproduct of writing code. A developer who wrote a change understood it because the work forced the learning. An agent removes that step. The developer can now submit a change they did not write and, in some cases, barely read. The thread's sharpest replies extend the point. One commenter, fwlr, argued the gate belongs earlier, before a developer asks for another person's time, not only before production. Another, rDr4g0n, described writing a full pull request description by hand, because describing a change in compact form requires understanding and often surfaces a hidden decision the agent made. The claim is not that agents write bad code. The claim is that shipping work you cannot explain transfers risk to reviewers, to users and to the next person who has to change it.
The AI verification burden in the numbers
The numbers behind that unease come from a working paper by Harvard researchers Fiona Chen and James Stratton, "Artificial Intelligence in the Firm: Bottlenecks in Software Production". The study used data from the engineering analytics platform Jellyfish, covering about 300 million work events from 725,938 workers at 718 firms between January 2021 and March 2026, as summarised in a report on the paper.
After firms adopted agentic coding tools, lines of code per worker rose about 30 percent, commits rose about 20 percent and pull requests rose about 23 percent. The resolution rate for tracked issues and completed epics, the study's proxy for shipped work, showed no statistically significant increase. Review absorbed the surplus. Time from pull request submission to merge rose 49 percent. The share of pull requests sent back for changes nearly doubled. Comments per pull request rose 35 percent. The share of workers doing code review rose 14 percent. The authors conclude that realising the benefits of AI adoption requires investment in downstream review capacity, not only in code generation.
The pattern is not limited to code. A 2026 Templafy report cited in coverage of the October agent launches found that 95 percent of professionals edit AI output, spending roughly four hours a week correcting it. Agents do not remove checking. They move it downstream, from drafting to verification.
What a quiz can and cannot prove
The Sphinx thread also contained the counter-argument, and it is correct in part. A developer can pass a comprehension quiz on code that is still wrong, because the quiz tests the developer's model of what the code should do, not whether the diff matches that model line by line. A separate analysis described the hook as "a level 1 control" and noted that its bypass, sphinx yolo, lets a push through but logs every skip. That logging matters. The gate is weak as a proof of correctness and stronger as a record of who accepted the risk and when.
That distinction applies beyond code. A check that only asks "did a person look at this" produces a signature. A check that asks "can this person explain what changed, what it depends on and what happens when it fails" produces understanding, or reveals its absence before the work reaches other people. Neither replaces tests, static analysis or review by a second person. Each solves a different failure.
Verification needs a record to check against
For knowledge work, the same burden is arriving without a Git hook to measure it. An agent can draft a summary, a policy note or an answer in seconds. The reader still has to decide whether the statement is true, current and based on a source they are allowed to see. That decision is slow when the underlying record is a pile of snippets retrieved by similarity, with no date, no owner and no link back to the decision that produced it. It is faster when the record itself is readable: a decision sits as an entry with the person who made it, the date and the sources behind it, and a draft waits for a person to review and publish it before it joins the shared record.
That is the design choice behind Jilu. The agent drafts. A person reviews and publishes. The published record stays open for any employee to read without asking the agent first, so verification is not a private chat with a model. It is a check against a shared, dated record that other people can inspect and correct. Agents will keep raising output. The firms that gain from that output will be the ones that built the review capacity, and the record, to absorb it.