Anthropic measured its own supervision bottleneck: agent hours outnumber employee hours six to one, and 80 percent of those hours are the agent waiting for a human
Anthropic published a study of its own workforce on October 16, and HN got to it on Sunday with 104 points and 188 comments. The authors are Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory, and the data is internal telemetry rather than a survey. The headline number: by August, interactive agent hours outnumbered employee hours six to one, and in March 2026 the sum of autonomous and interactive agent work hours passed human work hours for the first time. Code commits per engineer are running at five times the velocity of six months ago, measured with within-person fixed effects across 192 adopting engineers and 74 non-adopters, so it is not just the keen people being keen.
The part that matters for anyone running agents is the waste. Claude sessions amount to around 80 percent idle, waiting for human input. Even in non-technical groups, interactive sessions output enough text to fill over half of a work week if someone actually read it, at 0.75 words per token and 250 words per minute. Meta-work, meaning time spent writing instructions, prompts and standing rules for agents rather than doing the task, went from 10 percent of the average employee's interactive sessions in fall 2025 to 22 percent in spring 2026. Written standing instructions for agents grew from a few hundred in July 2025 to 8,500 a year later. The median employee spends 20 percent more time in meetings than a year ago. One quote in the paper does the summarizing: an army of Claudes, with every one asking for my attention. Agents show up in nearly half of the active public Slack channels, but the authors note they become teams for a single employee rather than nodes with high connectivity. The org-shape numbers are there too, with the junior share of technical hires down from 27 to 9 percent and technical departments' share of headcount from 61 to 56, and the company says it set up an early warning system for roles at risk; those will get discussed elsewhere. For this site the finding is simpler: the lab that sells the agents has measured that the agents spend most of their time waiting for the person who has to read them.
vLLM wrote the same finding from the other side on October 16. Its Kimi K3 post reports 840 tokens per second at batch one on 16 GB300s and 1.5 to 2x higher decode throughput, built with agent-assisted development of a unified CuTe DSL decode path. Then the sentence that belongs next to Anthropic's 80 percent: agent-optimized inference code is outpacing our review and maintenance workflow, so they need systems that allow automatic verification of reliability and performance. The answer is a fast-track model optimization RFC that isolates changes to model code and requires strictly better performance on all supported configurations before merge. That is the third major project in a week to say the same thing after the Claude Code team's review post and Postgres 19 slipping its release: generation is no longer the constraint, reading is, and the only fix that scales is to make the check mechanical.
Put the two together and you get the shape of the next year. Anthropic's number says a human attention slot is the scarce resource, and the agent sits idle for it 80 percent of the time. vLLM's RFC says the only way to free that slot is a verifier that does not need the human. Every product on this site's recent list, from Sphinx's understanding gate to Cursor's formal-verification skill to the HarnessΒ² and Doc2Harness papers, is an attempt to move work from the 80 percent idle column into something a machine can sign off. The lab's own telemetry is now the baseline against which those tools get measured. Microsoft and CMU announced a preceptorship program for early-career engineers on October 15, which is the training-side answer to the same chart.
Links: https://www.anthropic.com/research/employment-at-anthropic , https://news.ycombinator.com/item?id=50139873 , https://vllm.ai/blog/2026-10-16-kimi-k3-megakernel , https://www.cs.cmu.edu/news/2026/microsoft-cmu-praise
← Back to all articles
The part that matters for anyone running agents is the waste. Claude sessions amount to around 80 percent idle, waiting for human input. Even in non-technical groups, interactive sessions output enough text to fill over half of a work week if someone actually read it, at 0.75 words per token and 250 words per minute. Meta-work, meaning time spent writing instructions, prompts and standing rules for agents rather than doing the task, went from 10 percent of the average employee's interactive sessions in fall 2025 to 22 percent in spring 2026. Written standing instructions for agents grew from a few hundred in July 2025 to 8,500 a year later. The median employee spends 20 percent more time in meetings than a year ago. One quote in the paper does the summarizing: an army of Claudes, with every one asking for my attention. Agents show up in nearly half of the active public Slack channels, but the authors note they become teams for a single employee rather than nodes with high connectivity. The org-shape numbers are there too, with the junior share of technical hires down from 27 to 9 percent and technical departments' share of headcount from 61 to 56, and the company says it set up an early warning system for roles at risk; those will get discussed elsewhere. For this site the finding is simpler: the lab that sells the agents has measured that the agents spend most of their time waiting for the person who has to read them.
vLLM wrote the same finding from the other side on October 16. Its Kimi K3 post reports 840 tokens per second at batch one on 16 GB300s and 1.5 to 2x higher decode throughput, built with agent-assisted development of a unified CuTe DSL decode path. Then the sentence that belongs next to Anthropic's 80 percent: agent-optimized inference code is outpacing our review and maintenance workflow, so they need systems that allow automatic verification of reliability and performance. The answer is a fast-track model optimization RFC that isolates changes to model code and requires strictly better performance on all supported configurations before merge. That is the third major project in a week to say the same thing after the Claude Code team's review post and Postgres 19 slipping its release: generation is no longer the constraint, reading is, and the only fix that scales is to make the check mechanical.
Put the two together and you get the shape of the next year. Anthropic's number says a human attention slot is the scarce resource, and the agent sits idle for it 80 percent of the time. vLLM's RFC says the only way to free that slot is a verifier that does not need the human. Every product on this site's recent list, from Sphinx's understanding gate to Cursor's formal-verification skill to the HarnessΒ² and Doc2Harness papers, is an attempt to move work from the 80 percent idle column into something a machine can sign off. The lab's own telemetry is now the baseline against which those tools get measured. Microsoft and CMU announced a preceptorship program for early-career engineers on October 15, which is the training-side answer to the same chart.
Links: https://www.anthropic.com/research/employment-at-anthropic , https://news.ycombinator.com/item?id=50139873 , https://vllm.ai/blog/2026-10-16-kimi-k3-megakernel , https://www.cs.cmu.edu/news/2026/microsoft-cmu-praise
Comments