Human expertise stays load-bearing while the agent executes
Source → Observed → Interpretation → Model implication
Agentic coding and persistent returns to expertise
View source →Anthropic analyzed approximately 400,000 interactive Claude Code sessions from approximately 235,000 users between October 2025 and April 2026, using privacy-preserving classifiers built on Claude Sonnet 4.6.
Users make approximately 70% of planning decisions and Claude makes approximately 80% of execution decisions. A typical session runs about four turns, each prompt triggers about 10 Claude actions and occasionally more than 100, and Claude reads files, edits code, runs commands, and writes about 2,400 words of output per turn.
Novice sessions average about 5 Claude actions and 600 words per prompt while expert sessions average about 12 actions and 3,200 words. Anthropic reports the difference at p<0.001 and states that it remains significant when controlling for work mode, task value, month, occupation, and model family.
Verified success, measured from git commits and pull requests matching the work, passing test suites, and explicit confirmation, is 15% for novices and 28-33% for intermediate and expert users. At least partial success is 77% versus 91-92%. Most of the gain occurs from novice to intermediate and then flattens. Novices abandon troubled sessions at 19% against 5-7% for intermediate through expert users.
Domain knowledge matters more than coding background: non-software occupations reach 29% verified success on coding sessions against 34% for software occupations, and all ten largest occupation groups fall within seven percentage points.
Between October 2025 and April 2026 debugging fell from 33% to 19% of work, operating software rose from 14% to 21%, writing and analysis roughly doubled from about 10% to about 20%, and the estimated value of the average session rose 27%.
Anthropic excludes headless sessions run as a single prompt via the CLI and describes the excluded non-interactive usage as a substantial share of activity. It states that it cannot measure whether code written in a session is actually used or discarded thereafter, that classifiers remain challenging to validate at scale, that task value estimates are coarse, and that about 30% of sessions lack a classifiable occupation signal.
Apparition is documented at scale rather than asserted: the agent interprets intent, reads files, edits code, runs commands, and works through errors across chains of about 10 and occasionally more than 100 actions. The Selection content is that the variable separating useful outcomes from abandoned ones is the quality of human direction rather than the agent, and that this gradient did not decay across seven months of model change. Much of the context the agent needs still arrives through the operator, and the verification signal the study itself relies on is git and test evidence.
REFINES. The model places human attention in the denominator of agentic-system fitness, as a cost to be reduced. This evidence is consistent with human expertise also acting as an input that raises useful outcome: experts cause roughly double the agent activity per prompt and land verifiable results about twice as often, and the gap persists across model families and months. It also refines the Context condition, since a substantial share of the knowledge an agent uses is supplied by the operator rather than acquired by the agent. This does not falsify the Selection hypothesis; it suggests the human-dependency ceiling described by the 1K forcing function may bind through planning and context supply, not only through verification.
What this does not establish
- The source is Anthropic's own study of its own product and does not independently validate its usage, success, or value claims.
- It observes interactive sessions only. Headless and programmatic usage is explicitly excluded and described as substantial, so the source says nothing about orchestrated or non-conversational agent operation.
- It reports no concurrency measurement and establishes nothing about parallel agents, agent fleets, or 1K-agent scale.
- The expertise gradient is correlational. The source does not establish that expertise causes the difference in success.
- Verified success is a lower bound rather than a success rate, since the absence of a git or test signal is not evidence of failure, and Anthropic states it cannot observe whether the resulting code was used or discarded.
- Expertise and occupation are classifier-inferred, the source states that classifiers are challenging to validate at scale, and about 30% of sessions lack a classifiable occupation signal.
- Task value is described as coarse and not intended for literal dollar interpretation, so it does not establish an economics result.
- It does not establish the model's proposed evolutionary sequence.
Does the returns-to-expertise gradient survive concurrency? The measured unit is one person steering one agent, so the decisive next observation is whether expert direction still lifts outcomes when a single person supervises many concurrent or headless agents, or whether the supply of expert direction becomes the throughput ceiling the forcing function anticipates.