Brian Carpizo published a piece titled "Agile Is Dead. AI Killed It. Welcome Back, Waterfall." He nailed the diagnosis and picked the wrong destination. Agile is dying. Waterfall is not what comes next.
The diagnosis is right
Agile solved two problems that were real in the 1990s. Humans were bad at planning a big system up front, so a useful 200-page spec at the start was a fantasy. And building was slow and expensive, so being wrong at scale hurt. "Plan less, course-correct more" was the rational answer to both.
Both problems are gone. Not shrinking. Gone.
The planning problem broke when AI got good enough to sit in as a system-architecture partner. I can start with Claude at 9 a.m. and have a full architecture document by lunch: data model, API contracts, dependency map, eval rubric. The speed problem broke when prototypes got cheap. I have sat with a customer at 10 a.m. and shown them a working version of what they asked for at 11. Remove those two constraints and the ceremony on top, story points, planning poker, sprint reviews, the Scrum Master, the Product Owner as a buffer, is overhead nobody can justify anymore.
Why waterfall is the wrong answer
Waterfall did not fail because planning is bad. It failed because people could not hold a complex system in their heads, alone, in a slow-feedback loop, for months. That is the exact limit AI removes.
What replaces Agile is a spec-first loop. Write the spec with AI in hours. Treat it as context you feed the tools, not a contract you sign. Hand it to the AI as the primary input for code, tests, infra, and the customer brief. Ship a working prototype the same week. Put it in front of real users by week two. Keep the spec alive as you learn. Waterfall had nothing running for months. This has software running in days, with the architectural coherence Agile gave up on because planning used to hurt too much.
What dies, what lives
The ceremony dies. So does estimation theater, because story points were always a polite fiction, and the imprecision they papered over is the thing AI removes. The PM-as-translator role dies too: a senior engineer with Claude writes a stakeholder-ready spec in an afternoon. And "we'll figure it out next sprint" dies, because AI is literal. Hand it a half-formed thought and it builds you a half-formed system, fast, at scale.
Three things live. The customer feedback loop, because shipping to real people is still the only way you learn you were wrong. The empowered team, because AI multiplies each person, and four sharp builders beat twenty in a coordination org. And the bias toward working software, because a document with nothing running is a plan, not a deliverable.
The PM who gains here is the one who writes the system spec with AI, decides what to build, states it clearly enough for AI to build it, and owns the number that says whether it worked.
Try one thing this week. Pick the next real feature on your backlog. Before sprint planning, sit with Claude and write the actual spec: architecture, data model, API contracts, user journey, the metric you will watch. Give it half a day. Take that to engineering instead of the usual story breakdown. You will learn one of two things. The team ships 3x faster with a real spec, or your bottleneck was never story-point estimation. Both are worth knowing.