The Myth That an AI Agent Can Run Unattended All Day

Every new model launch sharpens the same promise: hand an AI agent a goal, walk away, and come back to finished work. Today's Claude Sonnet 5 release leans on exactly that image, a model cheap and capable enough to churn through long jobs on its own. The picture is seductive, and it is worth being precise about where it holds and where it quietly falls apart. The claim that an agent can run unsupervised for hours and land a complex task is, for now, more marketing than measurement.

The problem is not intelligence. It is arithmetic. An agent completing a long task chains dozens of steps together, and each step carries a small chance of going wrong. Those chances multiply. Researchers writing in a 2026 reliability framework for long-horizon agents put numbers on it: a 2 percent error rate per step compounds to a roughly 33 percent chance of failure across 20 dependent steps. Push the task longer and it gets worse in a hurry. One 2026 analysis of agents in production found that doubling a task's duration quadruples the failure rate rather than simply doubling it.

What actually goes wrong

The failures are not random. They cluster. Agents hallucinate the arguments they feed a tool, lose track of what they have already done, or fail to notice that a step did not work and barrel ahead anyway. Context is a big culprit. Even with a 200,000-token window, coherence starts fraying after 25 to 30 tool calls, at which point a model may forget an early result, redo a finished step, or contradict itself about the current state of the job. And because these agents take real actions, the mistakes are often irreversible. A wrong file deleted, an email sent, a payment made. There is no undo.

What the myth gets wrong, and what it gets right

The kernel of truth is real. Agents in 2026 genuinely handle work measured in hours, not the minutes they managed a year ago. The leap is real and the demos are not fake. What the myth strips out is the scaffolding that makes those hours reliable. The teams shipping dependable agents are not the ones with the newest model. They are the ones who break a big job into small, checkable pieces, build in explicit points to replan, and design a clean handoff back to a person when the agent hits the edge of what it can do. Anthropic's own research into 400,000 coding sessions pointed the same direction: the results came less from raw model horsepower than from the human steering the work.

So treat "runs unattended all day" as an aspiration, not a spec. The honest version of the pitch is narrower and more useful. An agent can carry a well-scoped task a long way before it needs you, and the skill worth building is knowing where that line sits. The human is not in the loop because the technology is primitive. The human is in the loop because, at a 2 percent error rate per step, that is the only place the work stays correct.