A platform migration used to be priced in engineer-weeks and breakage risk. With agents doing the work and adversarial review catching the drift, the bill is now my calendar time and an inference invoice. Almost every “not worth migrating” decision I made in the last decade was priced against a cost that no longer exists.
Author: James Phoenix | Date: September 2026
The Old Price of Moving
I have a couple of sites sitting on SiteGround. They have been there for years, not because SiteGround is the right home for them, but because moving was never worth it. That is the honest reason most software stays where it is. The hosting was fine, the platform was fine, and the cost of a rewrite was a month of evenings plus the real chance I would break URLs, lose redirects, drop images, and quietly bleed search traffic I would not notice for a quarter.
That cost had two parts. The first was labour: every template, every post, every redirect rule had to be touched by a human. The second was risk: the migration was only as good as my own attention, and my attention across a few hundred pages is not reliable. Migrations were expensive because the work was tedious and the verification was worse. The tedium I could grind through. The verification was what stopped me, because the failure mode was silent.
What the Bill Looks Like Now
I am uprooting those sites into Astro on Cloudflare, and the thing that made me finally do it is not that Astro is nicer or that Cloudflare is cheaper to run, although both are true. It is that the migration itself has become nearly free.
An agent crawls the old site, converts the content, rebuilds the templates, and writes the redirect map. Then a second pass, an adversarial reviewer, diffs the new site against the old one: every URL, every title, every canonical, every image, every internal link. The reviewer is not asked “does this look good?” It is asked “what did the first pass get wrong?” and it has the old site as ground truth to check against. Anything it finds goes back to the builder. Loop until the diff is empty.
What is left for me is deciding what the new site should be, reading the review findings, and paying the inference bill. The labour became inference and the risk became a review loop with an answer key. Neither of those is expensive. The whole thing costs less than the hosting renewal I was avoiding thinking about.
This is the same flip I described in The Harness Is Cheaper Now. There, the argument was that specs and tests stopped being overhead once implementation got cheap, because “figuring out what is wrong” became the bottleneck. A migration is that argument in miniature. The building is cheap. The review loop is the harness. Once the harness is cheap too, the only cost left is my time on the decisions the loop cannot make.
Why Migrations Are the Easy Case
I want to be precise about why this works so well here, because it does not generalise to everything.
A migration is fully specified by its source. The old site is the spec. Every page that exists on the old site should exist on the new one, at the same URL or behind a redirect, with the same content. That means the adversarial reviewer has something concrete to diff against, and the review can be mechanical rather than opinionated. As I argued in Answer Keys Are Authored, Not Found, a critic is only as good as the answer key it holds. A migration ships with its own answer key, which is why adversarial review is so brutally effective on it.
Greenfield work has no such key. If I ask an agent to build a new product and then ask a second agent to review it, the reviewer has no ground truth, only taste. That loop still helps, but it does not collapse the cost the same way, because nobody can tell me the diff is empty when there is nothing to diff against.
So the claim is narrower than “rewrites are free now.” It is: any transformation where the target is fully determined by the source has dropped to time plus inference. Migrations, framework ports, hosting moves, dependency upgrades, converting one storage format to another. The blast radius of a wrong answer is bounded, in the sense of the code review dial, because the reviewer can check every claim against something real.
What This Reprices
The consequence I keep noticing is that a large pile of past decisions were made against the old price and are now wrong.
“Don’t rewrite, it is never worth it” was good advice when a rewrite cost a quarter and had a silent failure mode. Under the new price it is often just inertia.
Hosting lock-in was a real thing when leaving meant weeks of work. Now the lock-in is whatever the agent cannot reach through an API or a crawl, which is much less than it used to be. The switching cost that protected a lot of platforms was labour, and labour is the part that fell.
Tolerating a bad platform “because migrating isn’t worth the hassle” is the one that stung. I stayed on shared hosting for years on that reasoning. The hassle was never the platform. It was the verification, and verification is exactly what got cheap.
The pattern is the same one I saw in The First Agents Were Human: the industries that converted first were the ones where the work was already checkable against a brief. Migrations are checkable against the source. That is why they are converting now, and why I expect the “never rewrite” instinct to look as dated in a few years as “never outsource the writing” looked to SEOs after 2023.
