Anthropic's 400,000 Claude Code sessions leave 70 percent of the planning decisions to the human

Anthropic's session data puts a length on the tightly supervised session: when users kept more than 80 percent of the execution decisions, Claude ran about eight actions per turn ^[6]. The same research says how much the agent finishes between check-ins tracks closely with how much control the user hands over ^[12].

The two figures being compared sit on different axes. Eight actions per turn describes sessions grouped by who made the execution calls, while the sixteen-action figure describes sessions where Claude made more than 80 percent of the planning decisions ^[7]. The rule the write-up draws from this is that you decide what and Claude decides how ^[8]. The longest runs in the data came from sessions where Claude decided both ^[2].

Sixteen against eight is twice the agent activity between check-ins, so a human looks up half as often for the same amount of agent work ^[1]. What that counts is how far the tool goes unattended. It is a real saving in interruptions for the person at the keyboard, and it is also a larger diff landing at review.

The research also says something about the operator. According to XDA's summary, the one thing a user cannot outsource is knowing enough about the problem they are trying to solve ^[9]. That sits next to the claim in the same piece that tools like Claude Code have lowered the barrier enough that years of programming experience are no longer needed to turn an idea into something functional ^[11]. Both can hold, and together they describe who the workflow is for: someone who can write a specification and grade the result, in a codebase they know well enough to say what done means.

What decides whether a team gets anything from the pattern is who writes the specification, and whether anyone on the team can tell a correct result from a plausible one. When the team writes the spec and can grade output, delegation works and the check-in saving is real. Write the spec but fail to grade, and you ship whatever compiles. Let the agent set the goal while keeping a strong reviewer, and the drift gets caught late, at review cost. Do neither and you have long sessions nobody can score.

The trainable artifact is the first prompt: what the outcome is, and what a finished result has to satisfy. Whether the training took shows up in how much of the agent's output a reviewer accepted without rework, and XDA's account of the research does not include that number.