LO
DeploymentCognitive automation2025-03-14
Airbnb reports migrating nearly 3,500 React component test files from Enzyme to React Testing Library in six weeks, against a hand estimate of 1.5 years
Frontend developeroccupation page →
Event date / reported
2025-03-14
Evidence stage
DeploymentAn employer has put it into production. Can move the baseline — weighted by scale and how similar the setting is.
Tasks this bears on
Owning the components everyone else uses
The shared library: deciding what belongs in it, what a change breaks, and who has to be told.
Still human-led✓ Evidence-backed
Where this applies
One company's own codebase, published by that company, with the hand-migration estimate (1.5 years of engineering time) supplied by the same company and not checkable from outside. What makes the result reproducible is the mechanism, which the post describes in full: the migration is a state machine whose every step has an automatic referee - jest, eslint and tsc say pass or fail - so the model could retry without a person watching, and the first bulk run cleared 75% of files in four hours. The remaining tail is the more useful number: after four days of tuning, 97%; the last 3% had each been retried between 50 and 100 times and were finished by hand. The post is also explicit that the main driver was selecting the right related files to put in the prompt - which grew to 40,000-100,000 tokens pulling in up to 50 files - rather than prompt wording. It reports no change to headcount, hiring or roles, and does not claim one.
What this means
The clearest published account of what an LLM migration of frontend code actually looks like, written by the team that ran it, with the failures in it. Two things generalise. First, the shape: it worked because jest, eslint and tsc can each say pass or fail without a person, so the machine could fail and retry thousands of times overnight — where a frontend task has an automatic referee, expect it to move. Second, the ceiling: three per cent of files survived fifty to a hundred automated attempts each and were finished by hand, which is a measured size for the part that did not go.
What it does not yet show
A refactor is not new interface work, and a test file has an oracle that a design does not. Nothing here bears on turning a mockup into something that renders, on the states nobody drew, or on what happens on a slow phone — and the post makes no claim about any of that. The 1.5-year hand estimate is Airbnb's own and was never tested against anyone actually doing it by hand, so the headline ratio is a comparison with a number the same company chose. And no role, headcount or hiring change is reported anywhere in it.
What you can check
Go through your own backlog and mark each item with what would tell you it is done: a green check in CI, or a person looking at it. The first list is the one that is about to get cheaper, and knowing its size in your own codebase is worth more than any industry figure.
Does it change the assessment?
No. The impact index is never moved by a single event. What this record did: the 1 linked task judgement above now rests on evidence instead of inference.
Source
The Airbnb Tech Blog (Airbnb's own engineering blog) · verified 2026-09-12 · Claude (VOLO agent) — post read end to end in the Browser pane (curl is blocked by Medium's Cloudflare interstitial, the script's own fetch is not); the 3.5K file count, the 1.5-year estimate, six weeks, 75%/4 hours, 97%/4 days, the hand-finished 3% and the 40k-100k token prompts are the post's own figures · interpreted 2026-09-12 · Claude (VOLO agent)
Primary source — published by the party that did this, or the authority of record. No co-signature needed.
This record is cited in