7 Hours Alone in a 12.5M-Line Codebase. 99.9% Correct.
Rakuten had Claude Code implement a method inside vLLM in one unbroken autonomous run. The headline isn't the seven hours — it's the 99.9% accuracy against a reference, because that's the number autonomy usually can't produce.
NeuroX AI · July 21, 2026
Anthropic's 2026 Agentic Coding Trends Report has one case that should reset your priors on how far a single agent run can go. At Rakuten, engineers pointed Claude Code at vLLM — 12.5 million lines across multiple languages — and asked it to implement a specific activation-vector extraction method. It finished the entire job in seven hours of autonomous work in a single run, hitting 99.9% numerical accuracy against the reference method.
Read that second number again, because it's the one that matters. Seven hours of unsupervised coding is easy to celebrate and easy to fake — an agent can churn for a shift and hand you plausible garbage. What separates this from a demo is that there was a reference: a known-correct implementation to measure against. The run wasn't graded on vibes. It was graded on 99.9%.
That's the whole game in production. Autonomy without a ground truth is just risk compounding by the hour. Autonomy pinned to a measurable target — a reference output, a rubric, a test oracle — is leverage. The agent didn't earn trust by running long. It earned it by being checkable.
The seven hours are the headline. The definition of "correct" is the system.