A deployment overhang, as Measuring AI agent autonomy in practice uses the term, is a situation in which the autonomy AI models are capable of handling exceeds what they exercise in practice — the latitude granted to agents in real deployments lags behind what they can handle.
Usage
Anthropic’s study draws the conclusion from two of its measurements of Claude Code. The 99.9th-percentile turn duration nearly doubled between October 2025 and January 2026, and rose smoothly across model releases rather than jumping with each launch, which the authors read as a sign that existing models are capable of more autonomy than they exercise. Among Anthropic’s internal users, the success rate on the most challenging tasks doubled between August and December while average human interventions per session fell from 5.4 to 3.3. The study contrasts its figures with external capability assessments such as METR’s, which measure what a model can do in an idealized setting with no human interaction, while its own measurements capture what happens in practice, where Claude pauses to ask questions and users interrupt. It stresses that the two kinds of metric are not directly comparable, and that together they suggest in-practice latitude lags behind capability.
In this account the gap is not only a matter of what users permit: the authors attribute the growth in autonomy to several possible factors, including power users building trust with the tool over time, applying it to more ambitious tasks, and improvements to the product itself.