Agents are waiting for us

Anthropic’s study of its own workforce reveals how productivity gains and declining demand for junior recruits can arise together.

Peter Stefanyi,Ph.D., MCC, Colaborix GmbH. October 2026

Two findings from one organisation

On 16 October 2026, Anthropic published How Employment is Changing at Anthropic. Drawing on employment records, agent telemetry, output metrics, surveys, calendars, Slack networks, hiring assessments and seventeen interviews, it offers an unusually detailed account of work in a company where AI agents have become central.

The report lends itself to two apparently opposing readings.

One begins with productivity. Code commits per engineer reached 5.1 times their level six months earlier, while contracts processed per legal employee roughly tripled: evidence that agents are delivering substantial gains.

The other begins with hiring. Juniors’ share of technical hires fell from 27% to 9% in about a year, suggesting that those gains accompany a narrowing route into employment.

A common mechanism may connect these findings. Across technical and non-technical roles, agents took over parts of implementation: writing code, finding bugs, logging progress and drafting emails. Automating this work can increase experienced employees’ output while reducing demand for people previously hired to perform it.

The question is how the organisation uses the released capacity and what happens to the people whose work changes. Anthropic makes both questions unusually concrete.

The wider adoption gap

The wider evidence reveals a gap between adoption and financial impact. McKinsey’s August 2026 survey found nearly nine in ten respondents reporting organisational AI use. Agent scaling reached 40% at large companies (annual revenue above $1 billion) and 22% at smaller ones. Although 80% reported personal productivity gains, only 37% attributed any enterprise EBIT contribution to AI. Just 6% both attributed at least 5% of earnings before interest and taxes to AI and reported significant value. This threshold describes a share of earnings attributed to AI, rather than investment returns or a five-point margin increase.

BCG’s September 2026 Applied AI Index offers a more encouraging, though differently defined, picture. Among more than 1,300 executives surveyed, 7.5% of companies fell into its most advanced category and another 41% were scaling AI and generating some value. Its results indicate that benefits are spreading, while McKinsey’s stricter earnings measure remains concentrated.

The distinction is between using AI, improving individual work and converting those improvements into organisational results. Anthropic is unusually well documented across that sequence. The employment study measures changes in output and work organisation; the external activity examined below provides a separate check on whether expansion is visible beyond internal throughput. Equity ownership and documented redeployment also show how employees’ incentives and changing roles were addressed alongside adoption. This combination makes the case more informative than an isolated productivity claim. It does not establish a net financial return or show that the transition benefits everyone, particularly prospective junior recruits.

The correspondence extends to organisational change. Nearly three-quarters of McKinsey’s high performers reported redesigning workflows, against about a quarter of others. They also pursued growth and innovation alongside efficiency. Anthropic makes such changes observable in detail, allowing their relationship to results to be examined.

What the numbers actually say

Output per person, indexed December 2025 = 100	Spring to June 2026
Commits per engineer	5.1×
Queries per data scientist	3.1×
Contracts per legal employee	3.1×
Deals per salesperson	1.7×
Hires per recruiter	1.4×
Invoices per accounts-payable processor	1.4×

The figures show substantial increases in output per person, but their interpretation depends on three distinctions.

First, the 5.1-fold increase is descriptive. Anthropic was also experiencing rapid business growth. To investigate the contribution of AI, the researchers followed 192 engineers who adopted coding agents and compared them with 74 who did not adopt during the study period, using fixed effects for individuals and calendar months. Code added rose sharply after adoption, although the confidence intervals were wide.

Second, output counts do not establish economic value. More code need not mean better software, as the research on writing versus shipping code cited by the report explores. Similar cautions apply to contracts, queries and invoices.

Third, the quality comparison is informative but incomplete. Anthropic compared agent-written and human-written code committed by the same employee, in the same month and directory, and found no difference in subsequent removal or reversion rates.

Output per person rose sharply as agents spread, and the adoption analysis supports an AI contribution. Its size and full quality consequences cannot be read from the headline multiple.

Anthropic develops and sells the technology it studies, and much of its work is software. Its employees hold equity and, in an August survey, expected AI to perform 75% of their work by 2030, against 28% among Americans generally. This intensive-adoption case is unrepresentative of the economy.

From internal output to external activity

When reading the research report I asked: Did the additional internal activity translate into externally visible work? Model launches, products and research publications provide evidence of expansion, though none directly measures labour productivity.

The following counts are my compilation of public releases and publications. They suggest faster and broader activity but depend on classification, especially where a launch could count as a feature or a new interface.

Model releases increased from six events in 2025 to eleven in the first three quarters of 2026: from 1.5 to 3.7 per quarter. Development depends on compute and capital as well as human work, so this cannot be attributed to agents or treated as a lower bound on their productivity effect.

Named product surfaces increased from ten in 2025, or 2.5 per quarter, to twenty-eight in the first three quarters of 2026, or 9.3 per quarter. A surface is a distinct named product, tool or interface. In 2025, the main adjacent product was Claude Code, previewed that February, alongside platform features and financial-services and life-sciences packages. The 2026 compilation spans more activities.

Claude surface launched in 2026	Category it enters
Cowork (Jan preview, Apr general release)	general office productivity
Interactive apps and connectors (Jan)	Zapier, integration middleware
Department plugins — banking, HR, PE, wealth (Feb)	vertical SaaS
Managed Agents (Apr)	agent-infrastructure startups
Claude Design (Apr)	Figma, Canva
Skills Marketplace, ~600 skills (May)	app marketplaces
Claude for Small Business (May)	SMB operations stack
Claude Security (May)	code-security vendors
Claude Science (Jun)	research workbenches
Claude Tag in Slack (Jun)	workflow bots
Claude Academy, Certified Architect (Mar, Aug)	training and certification
Claude for Financial Advisors (Sep)	wealth-management SaaS
Claude Docs and Claude Slides (Sep)	Google Workspace, Microsoft 365
Claude Dashboards and Claude Motion (Oct)	BI tools, motion graphics

This breadth matters: additional capacity had destinations beyond the original model offering. Launch counts establish neither commercial success nor agents’ share of the engineering effort.

Research expanded similarly: 44 publications in 2025 and 82 by 16 October 2026. For complete quarters, the figure records 71 publications in the first three quarters of 2026, or 23.7 per quarter, against 11 in 2025. The quarterly comparison uses 71; the mid-October count of 82 includes part of the fourth quarter.

The compilation records fourteen Science pieces after the Science Blog opened on 23 March 2026, covering protein design, analytical chemistry, biomolecular modelling, formal mathematics, quantum circuits and astronomy.

Sources and construction: commits per engineer come from the report’s smoothed Figure 3 series, indexed to December 2025 = 100 and ending in June 2026. External series are the author’s provisional compilation for complete quarters. See Sources for counting conventions and limitations.

Activity increased across these measures, accelerating around the end of 2025. Different units and observation periods prevent direct comparison of their magnitudes.

The 5.1× engineering figure also cannot represent the whole company. An illustrative calculation shows how including other functions changes the result, using assumed staffing weights and a convention for unmeasured output.

Function	Assumed share of headcount	Output multiple
Engineers	40% (assumed)	5.1×
Data science and legal	8% (assumed)	3.1×
Sales	10% (assumed)	1.7×
Recruiting and accounts payable	10% (assumed)	1.4×
Everyone else — nit measured	32% (assumed)	1.0× assumed
Illustrative weighted output index		≈ 2.9×

The report supplies the multiples, not the headcount weights. Technical departments account for 56% of employment, which constrains the context without establishing these weights. Holding other listed shares fixed and assigning the residual to unmeasured functions, changing engineering’s share from 30% to 45% moves the index from about 2.5× to 3.1×.

The central index is 2.9×, compared with compiled quarterly increases of 2.4× for model releases, about 2.2× for research publications and 3.7× for product surfaces. Their proximity does not validate the staffing assumptions: the external measures are related and share the same growing business context.

The calculation illustrates how other functions dilute an engineering headline. It does not estimate company-wide economic productivity: headcount weighting cannot make outputs of different value commensurable.

Higher throughput coincided with broader external activity. Separating the contributions of labour, agents, compute and capital requires evidence beyond these counts.

Revenue is the ultimate external indicator for a fast growing company and it speaks for itself revenue in Q1 of 2026 matched annual revenue of 2025 and kept growing as seen on the graph bellow.

The changing bottleneck

Organisations often adopt AI to relieve implementation constraints. At Anthropic, machine activity has created another: the human attention available to direct and evaluate it.

Agent working hours overtook estimated human hours in March 2026. By August, employee-directed interactive agents alone accumulated six times human hours; the research group’s ratio was about nineteen to one. Human hours assumed a forty-hour week; agent hours measured runtime, including inference and active tool use, excluding long waits. These measure activity rather than equivalent productive work.

Interactive agents also spent roughly 80% of elapsed session time waiting for human input, calculated by subtracting active work from total session time. This denominator differs from the working-hours comparison. Extensive machine activity still depended on human decisions.

The report estimates the time needed to read all agent output, assuming 0.75 words per token and 250 words read per minute.

Hours per person per week to read all agent output	April 2026
Engineering & research	424
Finance & accounting	43
Sales & go-to-market	35
Other non-engineering	27
Legal, policy & trust	26
Marketing & communications	12

Reading would absorb 30% of a forty-hour week in marketing, nearly two-thirds in legal, policy and trust, and more than the full week in finance and accounting. Engineering and research reach 424 hours, or 10.6 people reading full-time for one employee’s agents. These are hypothetical burdens, not observed review times; much output is code or intermediate material.

Exhaustive reading therefore cannot be the principal means of supervision. Verification can also use tests, automated checks, selective inspection and examination of the resulting system.

One engineer compared the attention demands to babysitting: a child needing a juice box opened can sound as urgent as one with a bleeding finger. Agents can likewise give trivial and consequential concerns similar prominence.

The scarce resource becomes the judgment to distinguish those cases and direct attention accordingly.

Work of greater scope and complexity

The output table counts activity per person: commits, queries, contracts, invoices, deals and hires. It captures throughput within functions, but less about changes in work’s difficulty and scope.

Elsewhere, the report shows employees attempting harder tasks and crossing established role boundaries.

An automated grader rated deidentified sessions by task depth, scope and the systems involved. After internal model releases, the share above the previous model’s 90th percentile of difficulty rose by around 30%, from 10% to 13%, remaining elevated for weeks. The easy end changed little. This measures difficulty attempted, rather than success on every harder task.

New capabilities extended the upper end of the work attempted while also accelerating familiar tasks.

The organisation of work changed alongside this expansion. Meta-work, meaning the construction of tools, specifications and monitors for agents, rose from 10% of the average employee’s interactive agent sessions in autumn 2025 to 22% in spring 2026, reaching about 29% in infrastructure. Written standing instructions grew from a few hundred in July 2025 to 8,500 a year later. In an April survey’s open-ended responses, 27% described productivity gains and 20% described a shift from doing the work to managing Claudes. These response categories need not be mutually exclusive.

The interviews give these changes substance.

A design team had written Claude.ai greetings such as “Hello night owl” and “Coffee and Claude time”. In spring 2026, it built a context graph describing the product, its voice and users, enabling agents to generate content. Having “scaled the team out”, as its manager put it, the writers moved into evaluations and prompts, applying their language skills to model behaviour.

A lawyer shifted towards frameworks that clients’ agents could discover: “SEO for legal advice”, making expertise accessible when needed. The report does not establish that legal judgment or accountability disappears.

Other examples are smaller but revealing. An engineer built an agent chief of staff to triage requests from other agents. Another wrote a performance review for their Claude team’s communication. A sales representative without an engineering background built a custom dashboard and sales pipeline in three or four hours. Managers’ code commits doubled. Among employees with at least a year at the company, the monthly likelihood of changing department was 50% higher in 2026 than in 2025.

Human coordination persisted alongside these changes. Agents appeared in nearly half of active public Slack channels. In the large department analysed in detail, human threads retained similar numbers of human responses as agent responses were added. Median meeting time increased by 20% over the preceding year, and by 16% among technical individual contributors. Company growth and changes in staffing contributed to the overall pattern. The findings nevertheless show that intensive agent use can coexist with substantial demand for human discussion and collective decisions.

The need for a durable organisational memory

Agents cannot update their assumptions from a meeting decision unless a record reaches them. One employee’s name remained on an old merchandise document, so agents kept directing t-shirt and mug requests to them after their responsibilities changed.

Documentation thus has immediate operational consequences. Current decisions and responsibilities help agents follow present intentions; stale records misdirect them. One researcher noted the cultural demand: leaders must defend feedback in the open. The report documents this pressure without quantifying transparency’s effect on output.

Organisational memory becomes infrastructure for agent work, dependent on accuracy and maintenance as well as coverage.

There is a sharper edge which we may call visible work. The report describes unlogged meetings and direct messages becoming invisible, confusing and discouraged. Recording decisions differs from opening every conversation to agents, although the demand for context can blur that boundary. Employers must determine what agents need, who may access it and where confidentiality belongs. How small can a durable record be while still supporting reliable action?

The missing juniors

Expanded roles for experienced staff accompanied weaker relative demand for junior technical recruits.

Juniors’ share of technical headcount fell from 21% to 14% in about a year; their share of hires fell from 27% to 9%, and postings to about 6%. Postings indicate future recruitment intentions.

These are shares in a growing workforce, not equivalent reductions in jobs. Anthropic tripled in size over the period, and technical hiring continued to increase in absolute terms. In non-technical departments, junior shares remained around a quarter of headcount and hires. The report describes internal redeployment, but these aggregate figures alone cannot establish every individual employment outcome.

Jobs using Claude for more complex tasks in late 2025 subsequently showed slower junior employment growth between January and July 2026. This association is consistent with reduced demand for junior implementation work, but establishes neither causation nor falling headcount in every such job.

A separate June study of roughly 400,000 Claude Code sessions found that the share devoted to fixing broken code fell from 33% to 19% between October 2025 and April 2026. That is a change in the composition of agent sessions, not a direct measure of the debugging work available to juniors. It is relevant because debugging has also been a route through which engineers learn why systems fail. The employment report raises the related concern that experienced employees are moving away from the object-level work through which they developed their skills.

Together, these findings suggest pressure on the reproduction of expertise. Practice tasks are being automated, recruitment favours experience and staff spend more time supervising. Neither study establishes subsequent capability decline. Erosion of a learning pathway remains an inference, with consequences that immediate output measures would miss.

Hiring assessments show a corresponding change in what the company values. References to candidates’ use of AI tools rose from around 5% of technical assessments in 2024 to roughly three times that share in 2026. In non-technical assessments, they rose from about 7% to roughly four times that share. For technical candidates, the largest increases were in comments on autonomy and judgment, while domain expertise occupied a smaller share of the assessment. For non-technical candidates, customer and mission orientation gained emphasis as execution skills lost share. These are changes in interview commentary, rather than a direct measurement of the contribution each skill makes to performance.

This need not contradict the June study’s finding that task-specific expertise predicts success with agents.

Expertise is being exercised differently. Verification still depends on knowing what correctness means. How will employees acquire and maintain that knowledge as agents perform more implementation?

What determines the employment outcome

As implementation is automated, human work shifts towards specification, coordination and judgment. Employment outcomes depend partly on where the released capacity goes.

At Anthropic, it coincided with expansion: twenty-eight compiled product surfaces in nine months, more research, harder tasks and greater internal mobility. The design team’s move into prompts and evaluations shows expertise finding a new use. Expansion appears to have enabled redeployment, without being its only possible explanation.

The authors similarly argue that prioritising internal reallocation encourages new tasks, products and markets beyond labour-cost savings. They contrast this with ‘so-so’ or inefficient automation: technology marginally outperforms displaced workers, delivering limited productivity gains alongside substantial social costs.

Expansion nevertheless does not separate winners neatly from job-cutters. Anthropic combines rising output with a narrowing junior entry route. Destinations for existing staff and opportunities for recruits are separate questions.

Anthropic recommends sharing financial benefits with employees, monitoring roles for disruption and prioritising internal redeployment. Rapid growth makes all three easier.

Against the wider gap between individual gains and organisational impact, four observations are particularly useful. Each links an organisational change to evidence in the report. They offer propositions to examine in practice rather than a validated sequence of instructions.

Reliable agent work appears to require a maintained record of decisions and responsibilities. The merchandise example shows why the currency and quality of context matter; more documentation alone would not solve the problem.

Human coordination remains substantial. Increased meeting time and persistent human Slack responses suggest that agents extend individual reach while difficult decisions still require collective judgment. These findings caution against treating discussion as an obvious source of savings, without implying that more meetings are inherently productive.

Greater capability changes what people attempt. The shift from 10% to 13% above the old model’s difficulty threshold indicates wider ambition alongside throughput. Sustainable work still depends on successful outcomes and demand, neither established by task difficulty alone.

Managing agents is itself substantial work: 20% of survey responses described this shift, meta-work reached 22% of interactive sessions and mobility increased. Specification, delegation and exception handling require expertise and time. The practical test is whether supervision enables useful work exceeding its cost while preserving the ability to recognise errors.

None resolves the report’s closing limitation: protections for incumbents cannot benefit those never hired. Anthropic documents destinations for existing employees, but leaves uncertain how newcomers will acquire the experience that work requires.

The agents are waiting for us. Supplying the judgment they require depends partly on developing it in the people who come next.

Sources

Wilmers, N., Catipon, R., Massenkoff, M. & McCrory, P. (2026). How Employment is Changing at Anthropic. The Anthropic Institute, 16 October 2026. All internal employment, telemetry, task-difficulty, survey, meta-work, meeting, mobility and hiring-scorecard figures, and all quoted employees.

Hitzig, Z., Massenkoff, M., Lyubich, E., Zhang, S., Heller, R. & McCrory, P. (2026). Agentic Coding and Persistent Returns to Expertise. Anthropic, 16 June 2026. Debugging share, expertise and verified-success figures, from roughly 400,000 Claude Code sessions.

On so-so automation and new tasks: Acemoglu, D. & Restrepo, P. (2019), Journal of Economic Perspectives 33(2); Beraja, M. & Zorzi, N. (2025), Review of Economic Studies 92(1). On writing versus shipping code: Demirer, M., Musolff, L. & Yang, L. (2026), NBER Working Paper 35275. All three are cited in the employment report.

External activity counts are retained from the author’s original compilation of model documentation, system cards, launch announcements, support release notes, the public timeline at github.com/jqueryscript/anthropic-claude-timeline and the research index at anthropic.com/research. Publication categories and product surfaces depend on classification. The weighted company index is an illustrative calculation using assumed staffing shares, not a finding of Anthropic’s study.

Primary sources: https://www.anthropic.com/research/employment-at-anthropic and https://www.anthropic.com/research/claude-code-expertise.

McKinsey & Company (25 August 2026). The state of AI in 2026: On the road to ROI. Survey: 1,719 respondents in 97 countries; self-reported results. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai

Boston Consulting Group (30 September 2026). The Formula for Agentic AI Value: The Applied AI Index 2026. Survey of more than 1,300 CxOs and senior leaders across more than twenty sectors. Its maturity and value categories differ from McKinsey’s EBIT-based definition. https://www.bcg.com/publications/2026/the-formula-for-agentic-ai-value