Agents Just Hit 66% — and 89% Still Never Ship
The 2026 Stanford AI Index shows agents leapt from 12% to 66% on OSWorld in a single year, within 6 points of human performance. Capability isn't the blocker anymore. 89% of enterprise agents still never reach production.
NeuroX AI · June 16, 2026
The 2026 Stanford AI Index just put a hard number on how fast agents got good: on the OSWorld computer-use benchmark, success jumped from 12% to 66.3% in a single year — within 6 points of human performance. On SWE-bench Verified, agents climbed from 60% to nearly 100% of the human baseline.
Then the same data drops the other shoe: 89% of enterprise AI agents never reach production. The model can drive a computer almost as well as you can, and nine out of ten agent projects still die in pilot.
So where do they die? A failure analysis of stalled deployments is blunt about it: 34% lose to scope creep, 27% to data quality — together 61% of all failures. Security blockers, integration mismatches, and 3–5x cost overruns take most of the rest. Not one of those is a model problem. They're the unglamorous engineering the demo skipped.
That's the whole gap in one sentence: capability is a solved curve, shipping is not. The teams that win in 2026 aren't the ones with the best benchmark score — they're the ones who locked scope, cleaned the data, and instrumented cost before the first agent ran.
We build that layer in 30 days, before your agent joins the 89%.