Here is a statistic the AI labs would love you to focus on. On a benchmark called OSWorld, which drops an AI agent into a real computer desktop and asks it to get things done, success rates leapt from around 12 percent to roughly 66 percent over the past year, according to Stanford’s 2026 AI Index. More than a fivefold jump in an agent’s ability to operate a computer the way you do. Terrifying, if you stop reading there.
What 12 and 66 percent actually mean
OSWorld tests agents doing ordinary desktop chores: opening apps, clicking around a spreadsheet, filing things in the right folder, stitching a few steps together. A year ago agents finished about one in eight of those tasks. Now they finish about two in three. That is a real and genuinely rapid improvement, and anyone telling you AI agents are standing still has not looked at the curve.
The boring catch
And yet, in the same week that number made the rounds, the actual reporting on agents at work told a duller, more useful story: coding agents still stumble over real tasks, the messy, underspecified, context-heavy work that makes up most jobs. The reason is not mysterious. A benchmark is a tidy, bounded test with a clear right answer. Your job is not. It is full of ambiguity, missing information, half-written instructions, judgment calls and the constant need to know when something has gone wrong. Two in three on a clean exam becomes something far less reliable once the task has no answer key, and a 66 percent success rate means the agent fails a third of the time, which for anything that matters is a lot.
That is the unglamorous truth under the scary headline. That gap is the whole distance between a demo and a colleague, and it is wide.
What this means
Agents are improving fast on the things that are easy to measure, and far more slowly on the things that are hard to measure, which happen to be the things your work is actually made of. Use them for the bounded, checkable chunks where a one-in-three failure rate is survivable, and keep a human in the loop for everything else. Your job is safe for now, not because AI is useless, but for the dull reason that real work refuses to sit still and be graded.
Sources: Stanford HAI 2026 AI Index; agent reporting, November 2026.