How employment is changing at Anthropic

What it claims

This October 2026 report from the Anthropic Institute measures how heavy agent use has changed work inside Anthropic. Its four main findings are fewer junior technical hires, higher output per person after agent adoption, human attention as the main bottleneck, and hiring criteria shifting toward autonomy and judgment. At five years old, Anthropic earns 2 to 4 times the revenue per employee that Meta and Google did at a similar stage of growth. The company also tripled in size over the past year.

How it works

The authors combine several internal sources:

• Agent session telemetry. Idle time is total wall-clock time minus the agent's active working time.
• Commit records. These support the output and code-quality comparisons.
• Reading-time estimate. Output tokens are converted to words and timed at average human reading speed.
• Task difficulty. An automated Claude grader scores deidentified sessions on depth, scope and number of systems involved.
• Hiring scorecards. An automated coder classifies interviewer comments for thousands of hires since 2024.
• Other sources. Calendar data, public Slack threads, employee surveys and 17 interviews.

To separate the effect of AI from company growth, the report compares employees in the same month who have and have not adopted agentic coding tools.

Evidence

Headcount and hiring

• Junior staff fell from 21% to 14% of technical departments since early 2025.
• Junior technical hires fell from 27% to 9%, and junior postings dropped sharply after December 2025.
• Technical departments fell from 61% to 56% of employees. Sales, finance and recruiting grew faster, and administrative headcount stayed low.
• In an April 2026 survey, about a third of ML managers and a fifth of SWE managers said they hire less because of AI.

Who writes code

• Coding-agent sessions by non-technical staff rose from 2 per month in August 2025 to 90 in May 2026.
• The share of technical employees not committing code fell by almost half in the year to April 2026.
• Managers' commits doubled, and the engineer to product manager ratio is about 17 to 1.

Mobility

• Employees with at least one year of tenure were 50% more likely to move departments each month in 2026 than in 2025.

Agent volume and output

• Combined autonomous and interactive agent hours first passed human hours in March 2026.
• Commits per engineer are 5 times their level of six months earlier.
• In the adoption study, 192 engineers who adopted coding agents added much more code afterward, compared with 74 who did not adopt.
• Code-quality check: agent-written and human-written code from the same employee, month and directory had no difference in later removal or revert rates.
• Sales representatives closed deals with more clients after adopting agents.

Effect of new models

• After each internal model release, tasks above the old model's 90th-percentile difficulty rose from 10% to 13% of sessions and stayed higher. The easy end of tasks changed little.

Attention and meta-work

• About 80% of interactive session time is agents waiting for human input.
• Standing agent instructions grew from a few hundred in July 2025 to 8,500 a year later.
• Meta-work (writing specs for multi-agent runs, agent-on-agent review, pipeline monitoring) rose from 10% of interactive sessions in fall 2025 to 22% in spring 2026.

Meetings and Slack

• Median meeting time rose 20% over the year, and 16% for technical individual contributors.
• Agents appear in nearly half of active public Slack channels. Human replies per thread did not change, and agents mostly post in dedicated Claude channels.

Surveys and hiring scorecards

• Agreement that work processes make staff "as productive as possible" rose from 44% in late 2024 to 58% in November 2025.
• Employees expect AI to do 75% of their work by 2030. Other Americans estimate 28%.
• Scorecard mentions of AI-tool use tripled for technical roles (from about 5%) and quadrupled for non-technical roles (from about 7%) between 2024 and 2026.
• For technical candidates, comments on autonomy and judgment grew most, and domain expertise shrank. For non-technical candidates, customer and mission orientation rose and execution fell.

Company responses

Anthropic has set up an early warning system that reads role-level demand and productivity metrics against AI capability and use. It flags jobs for internal redeployment. The authors also argue that broad stock ownership is needed to share the gains from adoption.

Limits

• Adoption is not random. Engineers who adopted fully autonomous agents were already on a steep upward output trend before adopting. Sales gains may partly reflect representatives with more opportunities choosing to automate.
• Output is not quality. The authors note that more code does not mean better code, and that AI errors may be harder to detect than human ones.
• Growth confounds the trends. Company growth overlaps with every trend, and the per-month adopter comparison only partly controls for it.
• The survey question changed. The April 2026 figure of 75% used weaker wording that dropped "as possible", so it is not comparable with earlier years.
• Some results are descriptive only. The link between slower junior employment and the most complex Claude use is not tested causally.
• Generalisation is limited. It is a single company, and the authors say many of its practices cannot be replicated across the economy.
• Junior hiring is unaddressed. The authors state that none of the company's responses address the decline in junior hiring.