One operator, by Anthropic's definitions: an agent-hours census

Abstract

Anthropic's How employment is changing at Anthropic (16 Oct 2026; https://www.anthropic.com/research/employment-at-anthropic) reports that by August interactive agent hours outnumbered employee hours six to one, and that interactive sessions sit around 80% idle. This note applies the report's definitions to one operator's local Claude Code and Codex transcripts. Across 9 full weeks the operator ran 4.0 interactive plus subagent agent hours per human hour, with 90% interactive idle share; headless runs were 6% of all agent hours. The four busiest weeks reach 5.0 to 6.8 : 1. Leverage runs 3.0 to 5.5 depending on whether subagents are counted and how human hours are measured. The escape rate of accepted agent output is not computed. One operator, two tools, 9 full weeks. The census script (standard-library Python; reads only; never prints prompt text) and every cited log are in the supplementary zip. The census script and this text were drafted with AI assistance (Claude Opus 5.5), and every figure was independently re-run before release.

2026 support ticket costs: OpenAI vs Anthropic prompt caching below $0.02 - cache or skip
Jonas Keller
Zenodo24 Oct 2026PreprintCC BY 4.0Full textZenodo
The Copernican Method Against the Copernican Principle: Anthropic Typicality over Partial Information States
Ajax Benander
Zenodo18 Oct 2026PreprintCC BY-NC-SA 4.0Full textZenodo
Falsifiable anthropic model from the string landscape
Zhiqi Huang
arXiv25 Sep 2026PreprintarXiv licenceFull textCosmology Gravitation & general relativityarXiv
Microbiological Quality and Analysis of Escherichia coli Resistance to Antibiotics in Beaches: Evaluating Anthropic Influence in Surface Waters in Brazil.
Monica M Elias, Francini Dos R Henrique, Philipe R Gomes +6
Water environment research : a research publication of the Water Environment Federation29 Oct 2026Peer-reviewedCC BYFull textEnvironmental science & pollutionEurope PMCPMC
On the foundations of anthropic reasoning: Implications for life across the universe
Brendon J. Brewer, Geraint F. Lewis, Jesse Goodman
arXiv16 Oct 2026PreprintarXiv licenceFull textInstrumentation & methods StatisticsarXiv

Cite

BibTeX RIS

BibTeX

@misc{condon2026operator,
  title = {{One operator, by Anthropic's definitions: an agent-hours census}},
  author = {Jared Condon},
  howpublished = {Zenodo},
  year = {2026},
  month = oct,
  doi = {10.5281/zenodo.23552652},
  url = {https://zenodo.org/records/23552652},
  note = {CC BY 4.0},
}

RIS

TY  - UNPB
TI  - One operator, by Anthropic's definitions: an agent-hours census
AU  - Jared Condon
PY  - 2026
DA  - 2026/10/22
PB  - Zenodo
DO  - 10.5281/zenodo.23552652
UR  - https://zenodo.org/records/23552652
AB  - Anthropic's How employment is changing at Anthropic (16 Oct 2026; https://www.anthropic.com/research/employment-at-anthropic) reports that by August interactive agent hours outnumbered employee hours six to one, and that interactive sessions sit around 80% idle. This note applies the report's definitions to one operator's local Claude Code and Codex transcripts. Across 9 full weeks the operator ran 4.0 interactive plus subagent agent hours per human hour, with 90% interactive idle share; headless runs were 6% of all agent hours. The four busiest weeks reach 5.0 to 6.8 : 1. Leverage runs 3.0 to 5.5 depending on whether subagents are counted and how human hours are measured. The escape rate of accepted agent output is not computed. One operator, two tools, 9 full weeks. The census script (standard-library Python; reads only; never prints prompt text) and every cited log are in the supplementary zip. The census script and this text were drafted with AI assistance (Claude Opus 5.5), and every figure was independently re-run before release.
ER  -

Abstract

Related

Cite