Anthropic reports AI agents spend 80% of session waiting for humans, remeasurement shows about 30%
Anthropic released a research report on October 16, 2026 that used internal telemetry, HR records, calendars and Slack data to quantify how AI agents are used inside the company. The report says AI agents work six times as many hours as humans and that, for interactive multi‑turn sessions, roughly 80% of wall‑clock time is spent waiting for human input. It also notes that meta‑work—writing…
Key points
- Anthropic's report says AI agents spend roughly 80% of interactive session time waiting for human input (as of August 2026).
- Independent remeasurement of 190 hours of Claude Code logs found about 30% waiting time, with 0.7% in self‑running sessions.
- AI agents' active runtime is six times human work hours, while meta‑work time rose from 10% to 22% of employee sessions.
The author re‑measured the waiting‑time claim using 190 hours of Claude Code logs from August 31 to October 19, 2026, covering 62,390 conversation events and 114 valid sessions. After applying Anthropic’s definition (excluding idle gaps over 30 minutes), the analysis found an overall waiting rate of 29.9% (median 28.3%), 31.5% in sessions with many human inputs, and only 0.7% in self‑running sessions. The author cautions that the differing measurement conditions—notably the scale of users and the definition of idle time—explain much of the gap between the 80% and 30% figures.
The piece situates these findings alongside other research, such as a METR randomized trial that showed developers were 19% slower with older AI tools, and a Stanford study reporting a slowdown in junior hiring in AI‑exposed roles. Together they suggest that while AI can generate output quickly, human inspection and instruction remain the primary bottlenecks, pointing to a need for better summarization and verification tools.
The story so far
2 episodes →- Anthropic reports AI agents spend 80% of session waiting for humans, remeasurement shows about 30%this story
80% of AI time is 'waiting for humans'—I re-measured Anthropic's findings against my own 190 hours of logs and found it was 30%
note.com · 19 October 2026
Loading the full article…
This text was published by note.com and written by yoshi. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- CoDA enables rapid codebook development for domain-specific text annotation · 1 src
- Researchers prove Prefix-Tuning can exactly steer pretrained transformers · 1 src
- Study probes if LMs understand variable roles via SAEs · 1 src
- OR-License benchmark tests LLM agents on real operations research tasks · 1 src
- Study finds most LLM agent decisions on prompt injections are set at read time · 1 src
Comments
via GitHub Discussions