Published Oct 16, 2026.
Key findings
- Anthropic has been hiring for a smaller share of junior technical positions as AI agent hours expanded; by August interactive AI agent hours outnumbered employees’ six to one.
- Output rose per person (code, sales, contracts) as employees adopted new AI tools; within-person adoption studies cited.
- Human attention is the bottleneck; reading all Claude agent output would take ~half the work week for many non-technical employees; rise in “meta-work” monitoring agent pipelines.
- Hiring managers increasingly emphasize autonomy and judgment over narrow technical implementation skills.
Labor demand
Junior technical share in technical departments fell from 21% to 14% since early 2025; junior hire share from 27% to 9%. Technical employment share of company fell 61% → 56% as sales/finance/recruiting grew faster.
Non-technical coding agent sessions (employees with no baseline commits) rose from 2/month (Aug 2025) to 90/month (May 2026). Engineers-to-PM ratio ~17:1 and rising.
Productivity
Agent work hours surpassed human hours company-wide in March 2026; by August employee-directed interactive agents worked 6× human hours. Code commits per engineer ~5× vs six months prior. Staggered adoption analysis on 192 engineers vs 74 non-adopters; code quality proxies (reverts, deletions) showed no clear AI “slop” gap in cited commit comparisons.
Sales reps closing more deals after agent adoption; lawyers, AP, recruiters cite automation examples.
Workplace dynamics
~80% of interactive session wall-clock time is idle waiting for humans. Standing agent instructions grew from hundreds (Jul 2025) to 8,500 a year later. Internal redeployment examples (product copy team → evals/prompting).
Anthropic tracks jobs at risk and prioritizes internal redeployments.
Methodology: internal employment, telemetry, surveys, calendars, hiring records, 17 interviews.