Anthropic reports that agent autonomy is shaped by models, users and products

Anthropic reports that agent autonomy is shaped by models, users and products

In a February 18, 2026 post, Anthropic reports on human-agent interactions involving Claude Code and its public API, asking how much autonomy people grant agents and how that changes with experience. Anthropic says the longest Claude Code sessions nearly doubled in duration over three months, while users with more experience both auto-approved more often and interrupted more often. It argues that this pattern reflects a shift from approving each action to monitoring and intervening when needed—not necessarily a retreat from oversight. Anthropic also reports that Claude Code asks for clarification more often as tasks grow more complex, while warning that the questions may be unnecessary or mistimed. For public API activity, Anthropic says most actions were low-risk and reversible, but usage was concentrated in software engineering and emerging in higher-stakes areas. It cautions that some apparently risky actions may be evaluations, and that it cannot see all downstream human review or reconstruct full API agent sessions. Its conclusion is that autonomy in practice is co-constructed by the model, user and product. Anthropic calls for post-deployment monitoring and tools that help people observe and intervene, while arguing against rules that require approval of every action. How well these findings generalize beyond Anthropic’s data and Claude Code remains unknown.

Publication: bc770d00-b524-2f58-3463-968bd1143efb

attribution — pass: The source is a single Anthropic research post making its own argument. The reading consistently attributes the findings, interpretations and recommendations to Anthropic rather than presenting them as established fact.

We analyzed millions of human-agent interactions across both Claude Code and our public API

summary — uncertain: The reading is faithful overall and keeps Anthropic's caveats and its argument against per-action approval mandates. However, central terms go unexplained for a general reader. It never says what an AI agent is (the source defines one as a system with tools to take actions), that Claude Code is Anthropic's coding agent, or what auto-approve means. Only API and red-teaming are defined. Smallest repair: add one plain clause each defining agent, Claude Code and auto-approve, or add them to the term library.

an agent is an AI system equipped with tools that allow it to take actions, like running code, calling external APIs, and sending messages to other agents.

date — pass: originalPublished is the instant 2026-02-18, and the reading states February 18, 2026, which matches. The three-month window and the October 2025 to January 2026 trend follow the source. Nothing is presented with hindsight. The page was edited after publication, but the cited passages are core body text consistent with the original post. This is a publisher page, so no archive disclosure is needed.

Between October 2025 and January 2026, the 99.9th percentile turn duration nearly doubled

support — uncertain: Most claims match the source. The phrase 'the longest Claude Code sessions nearly doubled in duration' (in the summary and claim 1) blurs what was measured. The source measures how long Claude works per turn before stopping, at the 99.9th percentile, not total session length. It also notes this extreme figure declined after mid-January. The page was also edited after publication (dateModified October 2026), though the cited text appears to be original. Smallest repair: say 'the longest stretches Claude Code works before stopping (99.9th percentile turn duration) nearly doubled'.

Among the longest-running sessions, the length of time Claude Code works before stopping has nearly doubled in three months

media — pass: No media was selected, and none was supplied or verified, so there is no media identity or relevance issue.