Are AI models held back from their autonomy potential?

Anthropic's own research report states that "Claude Code is working autonomously for longer." Looking at its longest-running sessions, Anthropic reports that "the length of time Claude Code works before stopping has nearly doubled in three months, from under 25 minutes to over 45 minutes." Measured precisely as the 99.9th percentile turn duration, the figure "nearly doubled, from under 25 minutes to over 45 minutes" between October 2025 and January 2026. Anthropic frames the tail of the distribution, not the average, as "the more revealing signal," since "the longest turns tell us the most about the most ambitious uses of Claude Code, and point to where autonomy is heading."

The mechanism Anthropic gives is that "this increase is smooth across model releases," which it says would not be expected "if autonomy were purely a function of model capability" — a sharp jump would track each launch instead. The smoothness instead points to several factors acting together: "power users building trust with the tool over time, applying Claude to increasingly ambitious tasks, and the product itself improving." Anthropic's stated conclusion is that "existing models are capable of more autonomy than they exercise in practice," and, combined with its separate capability evaluations, that the two together describe "a significant deployment overhang, where the autonomy models are capable of handling exceeds what they exercise in practice."

This measurement runs the opposite empirical direction from Does granting agents more autonomy undermine human oversight?, which argues current agent design erodes the human's capacity to oversee as autonomy grows. Anthropic's data instead shows people granting Claude Code autonomy more slowly than the model could absorb it — a gap on the human side of restraint, not (in this telling) a gap caused by eroded capacity. It also supplies something Can we design AI autonomy separately from capability? explicitly lacked: that paper's five-level taxonomy was offered with "no user study, deployment data" showing where real products sit on the scale; Anthropic's turn-duration trend is a concrete, if indirect, data point that actual usage is clustered well below the ceiling the models could support.

The measurement is specific to Claude Code and Anthropic's own API traffic, not to agent products generally, so it does not establish that other agents or domains show the same overhang. Anthropic itself cautions that agents are "difficult to study empirically" because there is no agreed definition of an agent and no reliable way to stitch independent API requests into sessions. The figures are Anthropic's own internal telemetry, not externally audited, and the explanation for the smooth trend ("several potential factors") is offered as a plausible reading rather than a tested cause.