Two stories from this week, from a database project and from a frontier lab, point at the same change: once a machine can produce or inspect work cheaply, the scarce resource becomes the human judgment that decides what to accept.
Postgres 19 slips, and the cause is better review
PostgreSQL 19's general availability, once slated for late September or early October, is now "expected to be generally available around November 19," and developers "dropped between eight and 10 significant features," The Register reported on October 16. Committer Peter Eisentraut (EDB) attributes much of it to LLM-assisted review. In his own October 9 post, he wrote that beyond a few features with "significant architectural defects," there was "a long tail of relatively harmless issues involving various edge cases" that "in past times... would not have surfaced until much later," to be fixed "bit by bit over the subsequent five years." Now: "with LLM-assisted code review, these kinds of things can get found much faster, and then you're staring at a list of like forty defect reports. And even if each of them requires only a three-line fix on average, there is a fixed overhead to dealing with each of these, and then you run out of time." He is pointed that this is not a "vibe coding" problem -- "the affected reverted features were written a long" time ago, before LLM review was useful.
Anthropic: human attention is the bottleneck
Anthropic's own October 16 employment report, drawing on internal data and 17 interviews, finds that "by August interactive AI agent hours outnumbered employees' 6 to 1," and that code commits per engineer rose to "5x their velocity as 6 months ago." The consequence is a bottleneck that has moved onto humans: Claude sessions spend "around 80% idly waiting for human input," and even for non-technical teams, interactive Claude sessions output enough to "take up over half of the work week." Employees describe shifting "from actually doing work to managing Claudes," one comparing it to babysitting: every agent "brings to you concerns with an equal demand for your attention." Notably, when Anthropic compared code written by the same person's agent against code they wrote directly, it found "no difference in rates of lines of code subsequently removed or reverted."
I think these two reports are the same finding at different scales, and the Postgres one is the more trustworthy because the Postgres team had nothing to sell. Eisentraut's number -- forty three-line fixes that each carry a fixed overhead -- is the whole story of this phase of AI-assisted engineering. The model did not create the bugs; it made the existing ones visible faster than a release schedule can absorb. That is genuinely good for software and genuinely disruptive to how teams plan, because the quality bar quietly rises mid-cycle. Anthropic's version is consistent but self-interested: a company that sells the agents reporting that its engineers are now bottlenecked by reading agent output is also an advertisement. The useful, falsifiable part is its finding of no difference in revert rates between agent-written and human-written code, which argues the extra output is not slop. My read for anyone managing this: budget reviewer hours as the real constraint, name one human who signs off on each deliverable, and treat a jump in found bugs as a reason to slow a release, not speed it.
What remains uncertain. Eisentraut himself says the cause is "probably a combination" of factors and may partly be coincidence; whether future Postgres cycles absorb LLM-found defects without slipping depends on whether review tooling keeps improving. Anthropic's data is from one unusually AI-saturated company and may not generalise, as the report's own "leading indicator" framing concedes.
Two October reports describe verification becoming the binding constraint as generation gets cheap: PostgreSQL 19 slipped to an expected November 19 and dropped 8-10 features after LLM-assisted review surfaced a long tail of fixable edge-case defects the beta had no time for (committer Peter Eisentraut says the code predates 'vibe coding'), while Anthropic's own employment study found agent hours outrunning staff 6:1, sessions 80% idle waiting on humans, and reading agent output eating over half the work week -- though it found no difference in revert rates between agent and human code.