On September 24, Ayman Nadeem published a short essay titled “Plan mode is dead.” It reached the front page of Hacker News and sat there collecting 339 points and 314 comments. His argument is that in Codex, “the boundary between planning and execution is disappearing as they collapse into one” — that the old loop of plan, approve, execute has become an unnecessary stopover, because the agent no longer needs the human to sign the itinerary.
He’s probably right. He is also describing something the comment thread spent 314 comments not noticing.
Plan mode was never a feature. It was a document.
The reason plan mode existed was not to make the model smarter. It was to make a human legible. Every plan → approve → execute cycle produced an artifact — a written statement of intent with a person’s attention attached to it — and that artifact is what a manager, an auditor, or a postmortem could later point at.
Nadeem’s complaint is that existing plan modes “didn’t feel collaborative enough.” Fair. But the collaboration was never the point; the paper was. Collapse planning into execution and the paper doesn’t get replaced. It gets relocated. Accountability reappears at the far end of the pipeline, in the pull request, where the person who approves a diff they never designed now owns the consequences of a design they never saw.
Northflank’s deployment write-up makes the shape of this clear: the enterprise review that actually gates an AI coding agent “does not ask which model scores highest on SWE-bench. It asks whether the agent survives contact with Okta, Splunk, and the code review policy.” Notice which of the three has no dashboard. Okta has an API. Splunk has a query language. The code review policy has a person with a browser tab open at 11:40 p.m.
Review is the only layer that doesn’t scale with tokens
Context Studios published its September 25 ranking of nine agents — Claude Code, Codex, Cursor, Copilot, Antigravity, Devin, OpenCode, Cline, Kiro — and it’s worth noting that Cline still ships a Plan/Act toggle. That toggle is now the industry’s appendix: harmless, vestigial, and evidence that something used to be there.
Meanwhile throughput goes up and reviewer headcount does not. Back in March, Tim Kellogg argued in “Plan Mode Is A Trap” that the ritual “sucks you into harmony with your agent without first setting ground rules” — and mentioned in passing that he’d run the same vibe-coding exercise past 10 to 15 job candidates. March is a geological era in this business, and the trap has only gotten more efficient since.
A staff engineer at a mid-sized logistics software vendor described the new arrangement to me in a pull-request thread that had, by that hour, attracted four approving emoji and zero substantive reviews: the agent had written 1,400 lines, and he had forty minutes before the release window closed. He approved it. That’s not negligence. That’s arithmetic.
The bill lands on whoever can’t refuse it
Here is the part that should make a certain kind of reader uncomfortable. For fifteen years, the right-of-center critique of enterprise software process was correct: change-advisory boards, sign-off theater, the committee that meets Thursday to bless a config change. Plan mode’s death is that critique winning. And the win is not free — it is simply invoiced to someone with less standing to dispute it.
The person who rubber-stamps a 1,400-line diff on a deadline is not exercising judgment. They are absorbing liability at the lowest available price. And it is usually the junior-most engineer in the room, which means the apprenticeship model that produced every senior engineer currently complaining about code quality is being dismantled in the same breath as the efficiency gain.
The apprentice used to write the plan. Now the apprentice approves the plan’s absence.
Keep the paper, lose the ritual
The answer is not to resurrect plan mode. Ritual without a document is just a meeting. But there is a real distinction between the ceremony of approving intent and the artifact of recorded intent, and we’re throwing out the second to be rid of the first.
Three things would survive the collapse. Plans that persist as reviewable artifacts even when execution starts immediately. Attribution per hunk, so that “who decided this” has an answer that isn’t “the agent.” And a review budget treated as fixed infrastructure — a real ceiling on merged lines per reviewer per day — rather than as a formality that will absorb whatever volume arrives.
Nadeem may be entirely correct that planning and doing are collapsing into one act. The open question is whether the human collapses into it too. The 314 comments suggest a lot of people believe that already happened, and are fine with it, because the tokens are cheap and the demo looks great.
The tokens were always the cheap part.