# wilmers-anthropic-employment-changing-2026-10-16

## Veille

Report from Anthropic's **Economics** team, authored by **Nathan Wilmers**, **Regina Catipon**, **Maxim Massenkoff**, and **Peter McCrory**, published on **October 16, 2026** on the Research site (approximately **4,500 words**, eight figures, PDF version). It studies Anthropic as a leading indicator of how work is organized under AI, drawing on internal data (headcount, agent telemetry, output, surveys, calendars, Slack, hiring interview rubrics) and **17 interviews**.

## Titre Article

How employment is changing at Anthropic

## Date

2026-10-16

## URL

https://www.anthropic.com/research/employment-at-anthropic

## Keywords

employment, labor market, Anthropic, junior roles, junior hiring, agent hours, AI agents, productivity, commits per engineer, staggered adoption, fixed effects, code quality, AI slop, human attention bottleneck, meta-work, agent management, standing instructions, meetings, Slack, observability of human work, hiring criteria, autonomy and judgment, internal redeployment, internal mobility, early-warning system, employee equity ownership, mediocre automation, so-so automation, postwar compromise

## Authors

Nathan Wilmers, Regina Catipon, Maxim Massenkoff, Peter McCrory — équipe Economics d'Anthropic ; rapport publié sur anthropic.com/research.

## Ton

Profile: a labor economists' report, in English, written in the first-person plural, published by the company under study using its own data. It addresses employers, public-policy decision-makers, and researchers.

Structure: key findings up front, then four parts (labor demand, productivity, bottlenecks and processes, what employers can do). Each part alternates a quantitative figure, its methodological note, and anonymous employee quotes.

Register: descriptive and cautious about the measurements. The limitations are written into the text: non-random adoption (selection bias), code quality measured via proxies, a survey question reworded between two waves, business growth that confounds the effects. The tone turns normative in the final section (*"Here we want to proceed carefully"*), which draws on history (railroads, the 1950s labor compromise) and the notion of *"so-so automation"*.

Imagery: babysitting (*"I need my juice box opened"*) for agents' demand for attention, *"an army of Claudes"*. The conclusion acknowledges that the proposed measures do not protect those who are not hired.

## Pense-betes

- **Source**: Anthropic Economics report, October 16, 2026, 4 authors, internal data + 17 interviews.
- **Revenue per employee**: at five years, **2 to 4 times** that of Meta and Google at the same stage.
- **Juniors**: share of technical departments **21% → 14%**; share of hires **27% → 9%**; job postings sharply down since December 2025. In April 2026, a third of ML managers and a fifth of SWE managers report hiring less because of AI.
- **Composition**: share of technical departments **61% → 56%**; engineer-to-PM ratio **~17 to 1**; manager commits ×2; administrative functions remained lean.
- **Non-technical staff who code**: coding-agent sessions **2/month (August 2025) → 90 (May 2026)**.
- **Mobility**: employees with one year or more of tenure are **50%** more likely to change departments each month in 2026; a product-writing team moved into writing evals and prompts.
- **Agent hours**: agents (autonomous + interactive) exceed humans as of March 2026; **6:1** in August for interactive sessions alone.
- **Productivity**: commits/engineer ×5 in six months; contracts, invoices, and sales up. Person × month fixed effects, treatment = 5 agentic sessions in the month.
- Quality: same employee, same month, same repository → no gap in deletion or revert rate at 90 days.
- **Difficulty**: after each model release, the share of tasks beyond the previous model's p90 rises from **10% to 13%**.
- **Attention**: **~80%** of session time = waiting on the human; reading all agent output would take more than **half a week** even for non-technical staff (0.75 word/token, 250 words/min).
- **Meta-work**: **10% → 22%** of sessions (multi-agent specs, cross-review, monitoring); standing instructions **a few hundred → 8,500** in one year.
- **Human**: median meeting time **+20%** (+16% among technical individual contributors); agents present in nearly half of public Slack channels, without reducing human responses.
- **Hiring**: mentions of creative AI tool use ×3 (technical) and ×4 (non-technical) between 2024 and 2026.
- **Expectations**: employees estimate AI will do **75%** of their work by 2030, versus **28%** for other Americans.
- **Levers**: equity ownership, occupation-level early warning, redeployment; public policy on retraining and income deemed necessary.
- ⚠️ **Scope**: the company's own data, non-random adoption; the authors note that no lever addresses the decline in junior hiring.

## RésuméDe400mots

On October 16, 2026, four economists from Anthropic's Economics team publish a report on the transformation of work within the company. They treat it as a leading indicator, because the company has rebuilt its workflows around AI. The report draws on internal data (headcount, agent telemetry, output, surveys, calendars, Slack, interview rubrics) and seventeen interviews.

Anthropic is growing fast but hires less per dollar of revenue than Meta or Google did at the same age. Junior technical roles are declining: their share of technical departments drops from 21% to 14% since early 2025, and their share of hires from 27% to 9%. Technical departments carry less weight in headcount, while non-technical staff start coding: their coding-agent sessions rise from 2 to 90 per month. Administrative functions remain lean, and internal mobility increases by half.

Since March 2026, agent hours have exceeded human hours; in August, employees' interactive sessions alone represent six times the human time. Commits per engineer have been multiplied by five in six months, and lawyers, accountants, and sales staff produce more. Comparing 192 engineers who adopt agents to 74 who do not, the authors observe a net increase after adoption, while acknowledging that adoption is not random. Code produced by an agent is neither deleted more nor reverted more often than code hand-written by the same employee.

Human attention becomes the bottleneck. About 80% of agent session time is spent waiting for a response, and reading all of their output would take more than half a week, even for non-technical staff. Employees respond with "meta-work": standing instructions, multi-agent specifications, cross-reviews, dashboards. This share rises from 10% to 22% of sessions. Meeting time increases by 20%, and work must stay visible so agents have context.

In hiring, evaluators mention creative AI use three to four times more often, and value autonomy and judgment more than domain expertise or execution.

The authors identify three levers: employee equity ownership, which shares the gains from adoption; an early-warning system for exposed occupations; and internal redeployment, which pushes toward creating tasks rather than automating at the margin. They acknowledge that none of these protects those who are not hired, and call for public policy on retraining and income support.

## GrapheDeConnaissance

- Anthropic —publie→ rapport How employment is changing at Anthropic (DOCUMENT, 0.98)
- Nathan Wilmers —travaille_chez→ Anthropic (ORGANISATION, 0.95)
- Maxim Massenkoff —travaille_chez→ Anthropic (ORGANISATION, 0.95)
- Peter McCrory —travaille_chez→ Anthropic (ORGANISATION, 0.95)
- Regina Catipon —travaille_chez→ Anthropic (ORGANISATION, 0.95)
- rapport How employment is changing at Anthropic —mesure→ part des recrutements juniors techniques passée de 27 % à 9 % depuis début 2025 (MESURE, 0.95)
- rapport How employment is changing at Anthropic —mesure→ heures des agents interactifs six fois supérieures aux heures humaines en août 2026 (MESURE, 0.93)
- rapport How employment is changing at Anthropic —mesure→ commits par ingénieur multipliés par cinq en six mois (MESURE, 0.92)
- Agents de codage —améliore→ production de code des ingénieurs adoptants, sans écart de suppression ou de revert à 90 jours (MESURE, 0.85)
- rapport How employment is changing at Anthropic —mesure→ environ 80 % du temps de session des agents passé à attendre une entrée humaine (MESURE, 0.9)
- rapport How employment is changing at Anthropic —affirme_que→ l'attention humaine est le goulot de la production des agents (AFFIRMATION, 0.92)
- Meta-work —observé_dans→ Anthropic (ORGANISATION, 0.92)
- rapport How employment is changing at Anthropic —mesure→ meta-work passé de 10 % à 22 % des sessions Claude interactives entre l'automne 2025 et le printemps 2026 (MESURE, 0.9)
- rapport How employment is changing at Anthropic —affirme_que→ les évaluations d'embauche valorisent davantage l'autonomie et le jugement que l'expertise de domaine (AFFIRMATION, 0.9)
- rapport How employment is changing at Anthropic —affirme_que→ le besoin de contexte des agents exige l'observabilité du travail et des décisions humaines (AFFIRMATION, 0.88)
- rapport How employment is changing at Anthropic —recommande→ Actionnariat salarié (CONCEPT, 0.88)
- rapport How employment is changing at Anthropic —recommande→ Système d'alerte précoce sur les métiers (METHODOLOGIE, 0.9)
- rapport How employment is changing at Anthropic —recommande→ Redéploiement interne (METHODOLOGIE, 0.9)
- Redéploiement interne —réduit→ So-so automation (CONCEPT, 0.78)
- rapport How employment is changing at Anthropic —s_inspire_de→ Compromis syndical d'après-guerre (CONCEPT, 0.8)
- rapport How employment is changing at Anthropic —affirme_que→ les protections des salariés en place ne bénéficient pas aux juniors non embauchés et doivent s'accompagner d'une politique publique (AFFIRMATION, 0.88)

---
Canonical: https://www.thekb.eu/en/fiches/wilmers-anthropic-employment-changing-2026-10-16/