How employment is changing at Anthropic

An Anthropic Institute report (October 2026) using Anthropic's internal employment, agent-telemetry, output, survey, calendar and hiring records, with 17 interviews, finds the junior share of technical hires falling from 27% to 9%, agent hours reaching six times employee hours, output per person rising across functions and after individual adoption of coding agents, and more time spent managing agents; the data are unreleased and the model-based coders unvalidated in the report.

Nathan Wilmers et al. (The Anthropic Institute) · 2026-10-16 · original

How does work change inside an organization that has rebuilt its workflows around AI agents? This report, published on 16 October 2026 by four authors on Anthropic's economics team, as a technical report of the Anthropic Institute (Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory), studies Anthropic itself, on the argument that a company "at the frontier of AI adoption" could be "a leading indicator" of changes in hiring, work organization and labor demand elsewhere (opening paragraph). It follows How AI is transforming work at Anthropic (December 2025), which surveyed and interviewed Anthropic engineers, and extends the question from self-reports to company records: headcount, hiring, agent telemetry, output metrics and calendars.

The report appeared on the same day as a companion survey, The AI expectations gap, and a short post presenting both, Working at the frontier shapes expectations for the future. It is a technical report by the developer about its own workforce and its own product, without peer review; none of the underlying data are released.

Version of record. The web page and a 20-page PDF ("Read in PDF") carry the same text; figure numbers below are the PDF's. In both versions, Figure 5's image is titled "Junior employment fell in jobs with high complexity Claude usage", while its caption reads "Junior employment growth by Claude usage complexity". The report has no methods appendix.

What the report does

The report combines several kinds of internal record, all described in the text and figure notes rather than in a methods section:

  • Employment and hiring records: headcount, hires and open roles by department and level, which give the junior share of each (Figure 1). "Technical departments" and "junior" are not defined beyond "lower band technical positions".
  • Agent telemetry: logged runtime of Claude agent sessions. "Agent hours" count agent runtime, "including inference and active tool use, but excluding any wait longer than 30 minutes", for employee-directed interactive sessions with subagents pooled; "human hours" are headcount times an assumed 40 hours a week (Figure 2 note). The ratio is therefore measured agent runtime against assumed, not logged, human time.
  • Output metrics for functions with countable output: code commits per engineer, queries per data scientist, contracts per legal employee, deals per salesperson, hires per recruiter, invoices per accounts-payable processor (Figure 3).
  • An adoption study of engineers and sales representatives. Each person counts as adopting in their first month with at least five agentic sessions; adopters are compared with not-yet-adopters using person and calendar-month fixed effects (a two-way fixed-effects event study), with 95% confidence intervals clustered on person (Figure 4 note). The engineer sample is 192 adopters and 74 engineers who did not adopt during the period ("Is AI driving productivity increases at Anthropic?"). The report does not give the number of sales representatives.
  • Model-based coding of text, called an "automated coder" or "automated Claude grader": of agent logs, to rate the complexity of each job's Claude tasks (Figure 5); of deidentified coding sessions, to rate task difficulty "based on depth, scope and number of systems implicated"; of a sample of 2,000 one-line chat summaries, to classify "meta-work" (Figure 7); and of hiring managers' interview scorecards "for thousands of employees hired at Anthropic since 2024". The report does not name the model or report any check of these codings against human judgment.
  • Internal employee surveys (late 2024, November 2025, April 2026, and a survey of managers in April 2026), calendar data, public Slack channels, and 17 employee interviews used to give context. Sample sizes and response rates for the surveys, and how interviewees were chosen, are not reported.

What it finds

Hiring and the composition of the workforce

  • Junior technical share. In technical departments, the junior share of headcount fell from 21% to 14% "since early 2025", and of hires from 27% to 9%; the open-roles line ends at 6% (Figure 1, which prints the end labels). In non-technical departments the final labels are 26% of hires, 25% of headcount and 13% of open roles. Postings for juniors fell "particularly sharply since December 2025" ("Labor demand at the frontier"). These are shares: the report gives no counts of junior hires or of total hires, and states that technical hiring "has continued growing in absolute terms", so a falling share does not by itself show fewer junior hires.
  • Timing as the report tells it. Discussion of lower demand for junior technical staff began in core technical functions in January 2026; by April, engineering and research teams had deemphasized hiring for "lower band technical positions". In an April survey of managers, a majority said they were hiring for different skills because of AI, and about a third of machine-learning managers and a fifth of software-engineering managers said they were hiring less because of AI ("Labor demand at the frontier").
  • Technical share of the company. Technical departments fell from 61% to 56% of employment over the year; technical hiring still grew in absolute terms, but hiring in sales, finance and recruiting grew faster. The report offers two readings: business functions expanding as the company scales, and coding tools letting non-technical staff build software.
  • Coding outside engineering. Coding-focused agent sessions by non-technical employees (no code commits at baseline, new hires excluded) rose from 2 a month in August 2025 to 90 in May 2026. Within technical departments, the share of employees with no commits in the prior 30 days fell "by almost half" over the year to April 2026; the engineer-to-product-manager ratio is about 17 to 1, and commits by managers doubled.
  • Movement between departments. Employees with at least a year at the company were 50% more likely to move departments each month in 2026 than in 2025. Administrative functions did not scale with the company's size, which the report attributes to agent assistants, automated review and outsourced ticket service.
  • Revenue per employee. At five years old, Anthropic's revenue per employee is two to four times that of Meta and Google at the same age (footnote 1: Anthropic's internal records for the twelve months to April 2026, against Google's and Meta's 10-K filings for 2003 and 2009, in 2025 dollars).

Agent hours and output

  • Agent hours. Counting autonomous and interactive agents, agent work hours first exceeded human work hours in March 2026. For interactive agents directed by individual employees alone, agent hours reached six times human hours by August 2026, and nineteen times in research organizations (Figure 2, which prints both multiples and uses a log scale).
  • Output per person. Indexed to December 2025, the latest plotted values are 5.1 times for commits per engineer, 3.1 times for queries per data scientist and for contracts per legal employee, 1.7 times for deals per salesperson, and 1.4 times for hires per recruiter and invoices per accounts-payable processor (Figure 3, a loess-smoothed index that prints these multiples). The report states that the company "saw explosive business growth" over the same period, so these trends alone do not separate the effect of AI from the effect of demand.
  • Within-person adoption. In the months after engineers adopt coding agents, the code they add rises sharply relative to non-adopters (Figure 4, "Engineers: code added"; the coefficients, which the figure plots without printed values, rise to roughly 1 to 2 log points in months 0 to 3, with wide intervals). Sales representatives close more deals with clients after adopting agents (Figure 4, "Sales: clients closed"). For fully autonomous agents, the report notes that adopters "were already on a steeply upward trajectory" before adoption, which "suggests" they were already increasing output.
  • Code quality. Comparing code written by the same person in the same month and directory, by an agent or by hand, the report finds "no difference in rates of lines of code subsequently removed or reverted". Figure 4's quality panel plots six proxies (lines removed within 90 days, removal by a bug fix, explicit reverts, lint suppressions, new TODO or FIXME notes, flaky-test markers); every plotted interval includes zero or, for explicit reverts, has its lower end at about zero (read from the figure, which prints no values; the unit is the AI-minus-human gap in percentage points). The authors add that AI errors "may be harder to identify than human ones".
  • Junior employment and task complexity. Across jobs, junior headcount growth from January to July 2026 is lower where Claude tasks were rated more complex in October to December 2025 (Figure 5, a weighted regression line through job-level points; no coefficient or interval is printed). The report introduces this pattern as descriptive and makes no causal claim for it.

Work processes and attention

  • Perceived productivity. An internal survey item on whether work processes made respondents "as productive as possible" drew 44% agreement in late 2024 and 58% in November 2025. In April 2026, 75% said yes, but the report notes "the question weakened by dropping 'as possible'", so the third figure is not comparable with the first two ("Bottlenecks and process").
  • Harder tasks after model releases. After internal model releases (the report does not say how many were studied), the share of interactive coding tasks above the old model's 90th percentile of difficulty rises from 10% to 13% ("around 30%") and stays higher in the following weeks; the easy end changes little. Difficulty is rated by the automated Claude grader.
  • Open-ended survey answers. In the April 2026 survey, 27% of answers to an open question on how AI had changed day-to-day work described productivity gains and 20% described a shift from doing work to managing Claudes. The report does not say who or what coded these answers.
  • Waiting for humans. About 80% of session time in interactive agent sessions is idle, waiting for human input (agent working time subtracted from wall-clock time).
  • Reading load. Converting each person's interactive output tokens to reading time (0.75 words per token, 250 words a minute), the latest values are 424 hours per person per week for engineering and research, and between 12 and 43 hours for the five non-engineering groups shown (Figure 6, which prints these labels). The report's text says that even for non-technical groups the output would take "over half of the work week". By this note's arithmetic, half of a 40-hour week is 20 hours; four of the five non-engineering panels (26 to 43 hours) are above it, and marketing and communications (12 hours) is below it.
  • Meta-work. Work about running agents (building Claude skills, writing plans and reviews for other agents, building dashboards and monitors, documenting how work is done) rose from 10% of the average employee's interactive sessions in August to October 2025 to 22% in April to June 2026; by function, the later share ranges from 12% (finance, legal and marketing) to 29% (infrastructure) (Figure 7).
  • Meetings and Slack. From calendar data, the median employee spends 20% more time in meetings than a year earlier, and technical individual contributors 16% more. Agents appeared in "nearly half of active, public channels"; in one large department's Slack network, agent threads added to human threads without reducing the number of human responses (Figure 8, a network drawing).
  • Hiring criteria. In the coded scorecards, mentions of a candidate's track record with AI tools appeared in about 5% of technical and 7% of non-technical assessments in 2024, and had "tripled" and "quadrupled" respectively by 2026. Comments on autonomy and judgment rose most for technical candidates, while domain expertise fell as a share; for non-technical candidates, customer and mission orientation rose and execution skills fell.

Interview material runs through these sections: an engineer describing "an army of Claudes, with every one asking for my attention"; an executive assistant who built a scheduling agent; a design team that moved to writing prompts and evaluations. These are illustrations chosen by the authors, not measurements.

The written instructions for agents

The report has no constitution or principles, but it measures one kind of written text that governs agents: "written standing instructions for agents grew from a few hundred in July 2025 to 8,500 a year later" ("Bottlenecks and process"). It also describes a product design team that, in spring 2026, built "a context graph that effectively guided agents to write compelling product content", and then moved to writing prompts and model evaluations; the report then quotes, without naming the speaker: "Writers who used language to drive human behavior are now using it to drive Claude behavior." The report counts instructions and reports this case; it does not test how any instruction's wording affects agent behavior.

What the authors propose

The final section ("What employers can do") is the authors' argument, not a measurement. They name three features of Anthropic's practice as worth considering elsewhere: broad employee stock ownership, compared with the post-war bargain in which unions accepted management control of work reorganization in exchange for pay tied to productivity; an "early warning system" of role-level demand and productivity metrics read against AI capability and use, which Anthropic "is standing up"; and a priority on internal redeployment. They state its costs: commitments to existing staff "could make employers hesitant to take a risk on hiring new employees", and "none directly address the slowdown in junior hires documented in this report". They call for government policy on retraining and income support alongside employers' actions.

Limitations and open questions

  • Causal identification. The authors state that adoption "is not random" and that people about to start a large project may adopt tools for it, which "would bias our estimates"; that sales representatives with more opportunities may select into automation; and that business growth confounds the output trends. The autonomous-agent panel shows a pre-trend. No instrument or randomized comparison is used.
  • What is measured. More code "does not necessarily mean better code or business impact". The quality proxies are deletions, reverts and markers, not reviewed defects, and AI errors may be harder to detect.
  • Model-based coding. Complexity, difficulty, meta-work and scorecard codes come from an unnamed Claude model; the report gives no validation against human coding.
  • Measurement choices. Agent-to-human ratios assume 40 human hours a week; reading-time estimates assume a conversion rate and reading speed; one survey item changed wording between waves.
  • Sampling and reporting. Survey sample sizes, response rates and the selection of the 17 interviewees are not given. The report is written by Anthropic economists about Anthropic, and the data are internal and unreleased, so outside replication is not possible.
  • Generality. The authors write that "many parts of Anthropic's employment practices" cannot be replicated across the economy. Anthropic tripled in size in the year studied and sells the product its staff use.

No outside analysis or replication of the report was found for this note.

Research questions

  1. How much of the post-adoption rise in code survives selection on timing? The authors flag that adoption may coincide with new projects. Using variation in when tools became available to a team, rather than when a person chose to adopt, would show how much of Figure 4's rise remains.
  2. Do the model-based codings agree with human coders? Meta-work, task difficulty and scorecard themes rest on an unnamed Claude coder. Human coding of a random sample of the 2,000 summaries and of the scorecards, with agreement statistics, would answer it.
  3. Is the fall in junior hiring specific to AI or to Anthropic's growth stage? The junior share fell while the company tripled. Comparing junior shares in other fast-growing firms with lower AI use over the same months would separate the two.
  4. Does agent-written code carry defects that the proxies miss? The authors note AI errors may be harder to identify. Audits of incidents and security findings traced to agent-written versus hand-written code would test the quality result.
  5. What share of the reading load is actually read? The reading-time figures are an upper bound on output volume. Logged reading or review time per session would show how attention is allocated in practice.
  6. Does the early-warning system predict which roles change? The system is described as being built. Publishing its flags and later role changes would show whether it works as intended.
  7. Do standing instructions change agent behavior, and which ones? The count of written instructions rose to 8,500. Comparing agent outcomes with and without particular instructions would show which matter.

How it connects

  • Earlier study of the same workplace. How AI is transforming work at Anthropic (December 2025), cited in this report, measured self-reported productivity gains among 132 surveyed engineers and researchers, with interviews and internal Claude Code transcripts; this report adds company records of output, hiring and time use, and describes its finding that commits per engineer rose fastest as consistent with that study ("Is AI driving productivity increases at Anthropic?").
  • Companion survey. The AI expectations gap, published the same day and cited here, reports that Anthropic staff expect AI to be able to do 75% of their work by 2030, against 28% for US workers; this report quotes that comparison. Working at the frontier shapes expectations for the future presents both reports.
  • Automating AI research. The report links When AI builds itself (June 2026) on the automation of model training and research.
  • Coding beyond engineers. The report links How Claude Code is used in practice (June 2026) when it discusses coding diffusing to other domain experts.
  • Retraining. Its policy paragraph links How well do job retraining programs work (August 2026), an evidence review co-authored by Massenkoff.
  • Junior hiring in the wider economy. This note connects the fall in junior technical hiring with Labor market impacts of AI: A new measure and early evidence (March 2026), which reported suggestive evidence of slower hiring of 22 to 25 year olds into AI-exposed occupations; the report does not cite it.

References