Nathan Wilmers, Regina Catipon, Maxim Massenkoff, Peter McCrory
How has AI restructured activity and occupation at Anthropic? As a business at the frontier of AI adoption, Anthropic could be a foremost indicator of how AI volition change activity organization, hiring, and labor petition throughout the economy.
Key findings
- Anthropic has been hiring for a smaller portion of younger specialized positions. This occurred as hours worked by AI agents related to workforce quickly expanded. By August interactive AI delegate hours outnumbered employees’ six to one.
- As AI agents diffused through the company, Anthropic began producing additional code per person, decision additional revenue per rep, and handling additional contracts. Following the identical employee complete time, output rosy as they adopted new AI tools.
- This acceleration makes individual notice the bottleneck for delegate output. Even among non-technical employees, study all the output of their Claude agents would obtain fractional their activity week. In response, workforce rotate to “meta-work,” of construction visibility tools and monitoring delegate pipelines.
- Changes in activity institution are reflected in shifting priorities among hiring managers, who increasingly emphasis autonomy and judgment, related to specialized domain ability and implementation.
Introduction
What transformations of activity and occupation can we anticipate under increasingly mighty AI? Anthropic is value studying since it has thoroughly molded its workflows about AI. Agents flit through Slack chats and agenda meetings. They autonomously fix bugs and collectively tackle huge engineering projects. Agent hours worked outnumber individual hours 6 to 1. Changes happening at Anthropic now provision advanced signals concerning a imaginable forthcoming of activity and employment.
This study draws on thorough inner Anthropic data covering employment, AI use, output metrics, employee surveys, inner communication patterns, employee calendar data, and hiring decision records. It contextualizes these quantitative data alongside 17 employee interviews.
We discover that equal as the business grows, related hiring in younger specialized positions has slowed. Productivity for the two engineers and non-technical workforce has leapt up and did so sharply whenever they adopted coding agents. Hiring managers increasingly emphasis applicants’ autonomy and judgement related to particular specialized expertise. A defining characteristic of operating in an AI-heavy institution is changeless force to streamline oneself out of this bottleneck stance whenever possible. One engineer described activity as facing “an armed forces of Claudes, alongside all one asking for my attention.” Employees at Anthropic remain critical. But the nature of their contributions has changed.
Anthropic has responded to these changes by environment up an first alert scheme to track jobs at hazard of disruption, and as jobs develop the business has prioritized inner redeployments. These actions, taken at the frontier of the two AI betterment and office change, display how particular organizational commitments can form the effect of AI on employment. But they additionally lift difficult questions concerning exactly what those commitments should be.
Labor petition at the frontier
Anthropic has expanded rapidly. But it has hired small per dollar of income than an before generation of ample tech companies. At five years old, Anthropic takes in between 2–4 times as much income per employee as Meta and Google as they crested in their own fast growth.1
What jobs are missing? Within Anthropic, conversation of a decrease in petition for younger specialized workforce emerged in center specialized functions in January 2026. The teams that relied most heavily on Claude in their own activity established that penning code and finding bugs could increasingly be offloaded to Claude. Newly capable models accelerated this perception: “the item that raised eyebrows was the uplift for achievement engineers in particular and it basically overnight changed what we were looking for.”
By April, engineering and investigation teams had deemphasized hiring for lesser collection specialized positions. When surveyed in the identical month, a bulk of managers described hiring for a distinct set of skills because of AI. Around a third of ML and a fifth of SWE managers reported hiring small because of AI.
The change in occupation shares is striking. Figure 1 shows that the portion of junior-level workforce in specialized departments at Anthropic slipped from 21% to 14% since first 2025. The proportion of younger specialized hires cut equal faster, from 27% to 9%. Postings for juniors have fallen particularly sharply since December 2025.
At the identical time, the general occupation portion of specialized departments inside Anthropic has fallen, from 61% to 56% in the final year. As the business expands, specialized hiring has continued expanding in complete terms, but intensive hiring in sales, finance, and recruiting has outpaced it. Part of this change is the natural enlargement of endeavor functions as the business scales. Part could indicate coding tools increasingly allowing non-technical workforce to create application themselves. For example, a revenue delegate designed a tradition dashboard and pipeline tailored to their work: “a build akin this was out of my wheelhouse, particularly alongside the UI, connecting all the tools, the MCPs. But I built this likely in three to four hours, fair iterating on it.” Coding-focused AI delegate sessions for non-technical workforce (identified as those alongside no code commits at baseline, excluding new hires) leapt from 2 per duration in August 2025 to 90 in May 2026.
Yet this change in hiring construction is not lone a narrative concerning automated coding allowing it to diffuse to other domain experts. Within specialized departments, the portion of workforce not actively coding (defined as those making no code commits in the previous 30 days) cut by nearly fractional complete the twelvemonth through April 2026. Part of this is because of a thinning portion of non-engineers: the proportion of engineers to merchandise managers, for example, sits at about 17 to 1 and has risen since 2025. The figure of commits from managers has additionally doubled. The supervising and coordinating roles inside specialized teams increasingly affect construction and shipping code.
One area in which non-technical hiring has not accelerated is in the administrative functions that typically measure alongside a company’s size. A blend of delegate assistants, automated assessment and ticket assistance outsourcing has remaining Anthropic clerical-light equal as it has grown rapidly. A financial director built a acquisition petition assessment delegate and explained to colleagues “it's fair trying to discover the things that obtain a lot of hours so they can concentration on the another stuff.”
Even clerical activity tightly connected to in-person communication has grown slowly. An administrator assistant, who built a individual aide delegate for scheduling and email triage, explained: “I think we've already moved former the item of Claude augmenting distant the really entry-level admin category role,” and alternatively focused their activity on higher-order “chief of staff” tasks.
These changes have already engaged reallocation: workforce operating at the business for at smallest a twelvemonth were 50% additional apt to move throughout departments all duration in 2026 than they were in 2025. Whole teams have additionally been reallocated as they handed off their activity to agents. Consider one merchandise scheme team, which wrote satisfied akin the greetings on Claude.ai: “Hello night owl;” “Coffee and Claude time?” In fountain 2026, the squad developed a environment chart that efficiently guided agents to compose compelling merchandise content. This “scaled the squad out,” its director explained, but additionally raised the inquiry of what was remaining for individual writers. The squad afterward rotated into delegate prompting and began penning example evals and prompts for Claude. “Writers who used tongue to run individual behavior are now using it to run Claude behavior.”
Is AI driving efficiency increases at Anthropic?
Are these shifts attributable to AI? At Anthropic, humans are a lean tier atop a expanding contingent of AI agents. Based on delegate meeting telemetry, in March 2026, the sum of autonomous and interactive delegate activity hours leapt former individual activity hours for the archetypal time. Part of this is fleets of agents conducting industrial-scale application engineering, filling in myriad regular tests and checks, conducting care on a expanding bulk of infrastructure code, and driving concerted campaigns to translate enormous codebases. But equal narrowing in on the agents deployed by idiosyncratic employees, as in Figure 2, by August the agents worked 6 times as many hours as their individual handlers.
Anthropic is a application producer, and the huge bulk of tokens flowing through these agents is spent producing application code. Dashboards detect fleets of agents functioning throughout the company. But individual aide agents additionally detect employee calendars and messages, scheduling, sending to-do lists accumulated from incoming messages, and automatically briefing workforce on the group and intent of their upcoming meetings. Tokens contact a expanding range of mundane activity tasks.
What do all of these tokens amount to? Figure 3 parcels output complete period for a range of job functions alongside measurable output. Consistent alongside previous reports surveying Anthropic engineers and describing automation of example training and research, code commits per engineer have risen the fastest, coming in at 5x their speed as six months ago. Data scientists are additionally querying the expanding bulk of Anthropic data more, asking Claude agents to get real-time metrics to dashboards and decisionmakers. But Figure 3 shows that equal non-technical workforce saw output increase as coding agents diffused: lawyers processed additional contracts; accounts payable ushered through invoices; salespeople closed additional deals.
But is all of this because of Claude usage? During this identical period, the business saw bomb endeavor growth. It is easier to near deals whenever you market a powerful model. To measure the autonomous consequence of AI, we difference the output of workforce who clasp agentic coding tools to those who have not yet adopted the tools. This lets us difference workforce operating in the identical duration at the company, and hence exposed to a akin endeavor context, but approaching it alongside distinct tools. Note, however, that tool acceptance is not random: if workforce on the cusp of driving several big new project decide to tool up for it, for example, this would bias our estimates.
We prosecute 192 engineers complete period as they clasp coding agents, compared to 74 engineers who do not clasp agents during the duration studied. Figure 4 shows that in the months following adoption, the amount of code that engineers add rises sharply. More code does not necessarily average improved code or business impact. To test for AI code slop, we study code commits made at Anthropic. We difference code committed by the identical employee, in the identical month, in the identical coding directory, but written by coding agents fairly than written immediately by the individual employee. There are many dimensions of code quality, and AI errors may be harder to acknowledge than individual ones. But we discover no difference in rates of lines of code subsequently removed or reverted.
Figure 4 additionally looks at what happens following an engineer adopts their archetypal completely autonomous agent. Those who clasp were already on a steeply upward trajectory compared to non-adopters. This suggests that workforce shifting toward completely autonomous agents were already ramping up their output.
Do these patterns go beyond coding, into a broader transformation of cognition work? Anthropic revenue representatives steadily adopted AI agents starting final September. Figure 4 shows that following revenue representatives clasp agents, they near deals alongside additional clients. Part of this form could be driven by representatives alongside additional revenue opportunities selecting into automation. But it could additionally indicate automation of logging and regular communication. One document administrator explained that Claude automations handled email example and client administration application integrations, replacing about 20% of their previous activity tasks. Another noted that akin automations freed them up to concentration on face-to-face time, letting them conversation to 30% additional of their customers than was earlier possible: “anytime you do item a second time, you're benevolent of like, okay, how can I use Claude to create it so that I don't have to do it a third time?”
Beyond sales, an Anthropic attorney built a sequence of Claude automations to scan risks and opportunities, triage lawful questions and obtain archetypal passes through public-facing copy for lawful review. A recruiter used Claude to assistance alongside applicant sourcing, tracking applications and an inspection that would have taken “maybe a brace of weeks,” freeing him to redesign outdated discussion questions and encounter alongside additional applicants.
All of this suggests that as AI capabilities improve, it impacts a broad range of jobs throughout Anthropic. Descriptively, we see younger occupation slowing in jobs that use Claude for the most complex tasks (Figure 5).
Bottlenecks and process: new and old tasks in an AI-intensive workplace
Heavy AI acceptance reshapes additional than occupation and productivity. It additionally drives continual changes in activity processes, generating new tasks and shifting employees’ perceptions of their workplace. One employee described this as the “eternal force at Anthropic to move as accelerated as imaginable and use Claude for everything” and another that “things we anticipate in a week are things that would obtain three months elsewhere.” An inner employee study asked whether activity processes at Anthropic made respondents as productive as possible: 44% of workforce accepted in delayed 2024; 58% by November 2025 (in April 2026, 75% answered yes, but the inquiry weakened by dropping “as possible”). As of August, employees at Anthropic expected that AI volition be capable to do 75% of their activity by 2030, compared to 28% for another Americans.
These ratcheting efficiency expectations are dramatized by the cadence of inner example releases. As shortly as a new example is available, Anthropic workforce commencement tackling harder problems in interactive coding delegate sessions. We measure this alongside an automated Claude grader ranking deidentified sessions’ project demands according to depth, range and figure of systems implicated. The portion of tasks complete the 90th percentile of difficulty for the old example rises sharply by about 30% (from 10% to 13% of all tasks) and stays elevated in the weeks following example releases. The uncomplicated end of tasks stated to agents changes little.
Spiraling delegate efficiency translates into output thrown throughout humans’ screens and attention. One engineer compared this to babysitting:
They're like, I need my beverage box opened and it's an emergency that my beverage box is opened. And that is voiced alongside the identical flat of involvement as the one who has sliced their digit and is bleeding. [...] When you are babysitting, all kid brings to you concerns alongside an equal petition for your notice and the assorted [Claudes] additionally do that.
In the April 2026 survey, an open-ended inquiry concerning how AI has affected day-to-day work, 27% described efficiency gains from Claude tools and 20% described their function shifting from really doing activity to managing Claudes.
Amid this model-driven acceleration, individual notice is a changeless bottleneck. In interactive, multi-turn sessions, the huge bulk of agents’ meeting period is not spent operating at all, but spent waiting. We track idle period by subtracting all agent’s energetic operating period from all interactive session’s total wall-clock time. Claude sessions amount to about 80% idly waiting for individual input. As one engineer put it, group are “bottlenecking the agents.”
How much period exactly would be needed to peruse all of Claude’s output from interactive sessions? For all interactive session, we change all session’s delegate output token to term count, assuming average individual study speed. Figure 6 shows that specialized functions have seen the quantity of coding delegate output jump up complete the final six months. But equal for non-technical groups, interactive Claude sessions are outputting adequate satisfied to obtain up complete fractional of the activity week. The deluge of Claude content needs organizing.
One resolution to this deluge is rising automation. But this requires activity of a distinct kind. Written position instructions for agents grew from a few hundred in July 2025 to 8,500 a twelvemonth later. One engineer set up an delegate chief of personnel to triage delegate requests. Another wrote a achievement assessment for their Claude team: “Hey, let’s conversation concerning what is and is not operating concerning our communication style.” Recruiters built unified dashboards alongside agents surfacing problems and initiating actions. A attorney shifted their notice from reviewing lawful questions to doing “SEO for lawful advice, you need to compose frameworks that [clients’] agents discover.”
Figure 7 shows that this benevolent of meta-work leapt from 10% of the average employee’s interactive Claude sessions in autumn 2025 to 22% in fountain 2026. Much of this increase was due specifically to delegate management: penning specs for multi-agent runs, looping one delegate to assessment another’s output, and monitoring pipelines. Across functions, individual involvement increasingly distracted from a particular part of code or penning to the tooling necessary to get agents to do their tasks effectively.
Despite increases in delegate reliability and these important investments in tooling to supervise them, individual rely and accountability remains critical. One employee explained that, increasingly, chasing downward bugs “uncovers these organizational or systemic or tech debt things that really necessitate several amount of individual convening and decision-making and governance to solve.” In discussion following interview, workforce described face-to-face coordination as indispensable. Based on calendar data, the median employee spends 20% additional period in meetings than a twelvemonth ago. Part of this reflects the growth of the company, and a change toward meeting-intensive non-technical roles. But equal among workforce in specialized idiosyncratic contributor roles, gathering period is up 16% complete the former year. Big-picture strategic questions contact a broad range of leveraged positions, and typically necessitate conversation and involvement alongside others. A interrupt from managing the agents typically looks akin jumping into a meeting.
Broadly, the two specialized and non-technical roles have seen implementation parts of their job automated (writing code, identifying bugs, logging progress, example emails). Engineers described becoming leaders and managers of agents; non-technical workforce described doing additional and additional of the strategic parts of their job: “that 20% that was strategic becomes 90% of the job.” But for many non-technical roles, center implementation activity is additionally interpersonal: a revenue call, an interview, cross-functional alignment on a complex problem. All of this remains human.
Consider too the Slack channels that provision the two the forefront leaf and backchannels of Anthropic communication. Agents diffused through Slack and eventually appeared in nearly fractional of active, community channels. Figure 8 zooms in on the communication network of a sole ample department, example on all community Slack channels that division members interacted through. It links the two individual and delegate participants by whether they participated in a Slack thread. The fig shows the notable increase of agents, as orange dots blossom and dominate the network.
Agents appear disproportionately in dedicated channels for Claude-human interactions, fairly than messaging a lot in incumbent individual channels. The underlying construction and density of individual communication changes little. Agents did not substitute individual interaction, but fairly added to it: individual threads in this division sustained the identical figure of individual responses, during additionally layering on delegate responses. Agents rotate into teams for a sole employee, fairly than nodes alongside elevated connectivity to many workforce at once.
As humans activity amid a expanding fleet of AI agents, they visage force to change individual activity into agent-visible channels. If a decision is made in an in-person meeting, and no notes or report-out appears, agents are missing crucial environment for updating their goals and assumptions. An employee explained how having their name on a months-old document concerning Claude swag meant agents kept bringing requests for t-shirts and mugs, lengthy following their responsibilities had changed. Private work, in unlogged meetings and straightforward messages, is invisible, confusing and discouraged in an AI-heavy workplace. One investigator described this default toward openness as crucial to Anthropic’s thoroughgoing integration of AI: “most organizations do not run that way and I think that's a big change since leaders have to defend feedback out in the open.” The need to nourish AI environment requires observability of individual activity and decisions.
Who volition thrive in a office akin this? Shifts in the satisfied of activity at Anthropic additionally display up as changes in the skills that hiring managers prioritize in hiring decisions. We used an automated coder to procedure hiring director discussion mark cards for thousands of workforce hired at Anthropic since 2024. These comments are qualitative reactions by interviewers assessing candidates’ performance. The clearest alter is risen accent on whether candidates have demonstrated imaginative use of AI tools. In 2024, hiring managers noted candidates' track document on AI tooling in about 5% of assessments for specialized roles, and 7% for non-technical roles. By 2026, these rates had tripled for specialized roles and quadrupled for non-technical roles. Hiring managers increasingly concentration on whether candidates efficiently use AI.
Deeper changes in appraisal are additionally visible. Among interviews for specialized candidates, the sole biggest increases were in comments on applicants’ autonomy and judgment. Discussion of domain ability shrank as a portion of general assessment. As one director explained, “I anticipate models to continue to be jagged, so for my squad I desire them to be fungible throughout domains.” Similarly, for non-technical candidates, client and goal inclination increased, during conversation of implementation skills fell. Increasingly, hiring managers measure applicants according to their judgement and commitment, fairly than the domain-specific skills and implementation automated by AI tools.
What employers can do
The profound acceptance of AI at Anthropic makes it a helpful case. As AI diffuses and improves, the institution of activity could change as it has in previous waves of scientific change. The railroads gave us the managerial corporation, steam power allowed factories, and individual computers delivered the startup. What new organizations volition tool-using, reasoning LLMs require, and what volition be our function in them? One possible is single-person startups, all alongside a individual originator steering swarms of agents to rake in billions in revenue. Another is companies built by AI brilliant CEOs employing individual agents lone to wield pipettes and attest before Congress.
Anthropic does not appearance akin either of these. It has tripled in size in the final year. It is no longer a small startup. It has additionally not ceded administrator authority to its models. Written Slack debates and in-person meetings steer everything from business scheme to squad decisions, alongside small input from AI agents. Keeping individual activity apparent throughout the business ensures autonomous agents logic from caller context. Individual contributors orchestrating teams of agents appearance additional akin managers in conditions of the cognition they need and the interdependencies in their work. Working and deciding in the open is crucial at an AI-intensive organization.
As workforce increasingly see their activity as managing teams of Claudes, they expend additional period construction the tools needed to do this efficiently in their own distinctive domains. This experimentation method that workforce are continually automating their own tasks away. In several roles, they are increasingly removed from the object-level activity that initially built their skills. As agents sprint through workflows, individual notice and judgement are a bottleneck that the scheme strains against. From revenue to legal, individual workforce are expected to do additional and distinct tasks. Roles broaden to contain construction products, managing agents and tackling the strategic and interpersonal activity that remain beyond the range of the finest LLMs.
These changes in the institution and satisfied of activity have appeared alongside declining shares of younger specialized roles in headcount, hiring, and posting. Hiring managers increasingly concentration on applicants’ autonomy and judgment, fairly than implementation and narrow expertise. Overall, Anthropic’s income growth related to its occupation growth has sped former equal the fast first efficiency growth of a previous generation of innovation companies.
Reading these patterns onward and out throughout the economics suggests that AI diffusion could hazard employee displacement. Prior activity additionally finds that AI-native startups and AI-adopters are flatter and employ fewer entry-level employees. What can we study from Anthropic concerning how to oversee AI upheavals in a way that serves the humans operating about them?
Here we desire to continue carefully. There are many parts of Anthropic’s occupation practices that cannot be replicated throughout the economy.
But three features remain out. First, general inventory ownership gives workforce a powerful involvement in the company’s success. Of course, this is an uncomplicated posture for a business that has seen significant increases in valuation. For many firms, shared ownership may be insufficient or alternatively undesirable. But it is difficult to ideate a type of pro-worker AI acceptance in which workers do not clasp a financial interest in prosperous rollout. If idiosyncratic contributors are expected to “manage agents,” “make VP-level decisions” and be “their own endeavor line,” as interviewees put it, afterward aligning their interests alongside the institution is crucial. The post-World War II union-management colony forged in the 1950s rested on exactly this benevolent of bargain: labor unions akin the United Auto Workers and the United Steel Workers allowed administration prerogative complete activity reorganization (and equal intensification) in come back for pay increases tied to efficiency increases and to a affluent gathering of benefits. Whatever the new agreement between AI-affected workforce and their employers, a scheme to portion benefits of acceptance is essential. At Anthropic, workers codify their ability for the machine, see their activity routines continually roiled and disrupted, and have seen a ample part of their job rotate into speedreading the output of their AI collaborators. They do not do this for free.
Second, Anthropic is position up an ongoing first alert scheme to detect jobs facing AI-driven transformation. A sequence of role-level metrics for petition and efficiency are peruse against AI capabilities and use. These total metrics assistance emblem anywhere activity is shifting. An first alert scheme allows the business to commencement reallocating affected workforce to jobs alongside a skills fit that may be small immediately at risk. As companies throughout the economics reorient to an accelerated gait of activity reorganization, productive systems for directing inner employee reallocation volition rotate into increasingly valuable. Turning to the external labor market to encounter all new project demanded by AI advances volition be a stronger bottleneck the faster AI advances. In this report, we established risen movement between departments, changing activity content, and workforce shifting to new tasks as old ones were automated. Building systems to flexibly reallocate workers internally volition be crucial for AI-intensive organizations.
This dedication to inner employee reallocation raises crucial tradeoffs. A well-functioning labor market should change workers into their most productive positions, whether that is at their current director or another one. Commitments to reallocate incumbent workers could create employers hesitant to obtain a hazard on hiring new employees. In normal financial conditions, particularly intensely rivalrous ones, specified commitments could be costly.
But prioritizing reallocation forces AI-adopting companies to regard ways that AI can create new tasks for new products and new markets, fairly than funneling AI applications into replacing individual workers to preserve on labor costs. This could assistance employers oppose the enticement of ‘so-so’ or ‘inefficient’ automation in which AI or another technologies fair barely outcompete individual workers: employers substitute individual workers, imposing significant social expenses in misplaced earnings for displaced workers, during delivering minimal efficiency benefits.
From what we see in Anthropic, the two novel and persistently individual tasks remain alongside intensive AI use. Uncertainty concerning the forthcoming set of these tasks should propose preserving a elastic workforce. As one Anthropic administrator noted, “a workforce that is anxiously concerned is not a workforce doing its finest work.”
None of these actions volition forestall mighty AI from disrupting jobs. And certainly none immediately location the slowdown in younger hires documented in this report. Employment protections for incumbents volition not advantage those not hired to commencement with. Private act on the part of employers must arrive alongside farsighted authorities guideline about retraining and earnings support. Such guideline have to be guided by a apparent image of the measure of office transformation that AI can bring.
Suggested citation
@techreport{wilmers2026employment, author = {Nathan Wilmers and Regina Catipon and Maxim Massenkoff and Peter McCrory}, title = {How occupation is changing at Anthropic}, institution = {The Anthropic Institute}, type = {Technical Report}, year = {2026}, month = oct, url = {https://www.anthropic.com/research/employment-at-anthropic} }Acknowledgements
We are grateful to the Anthropic workforce who gave period for interviews concerning how their activity has changed. For extremely helpful comments and assistance, we appreciate Kate Astbury, Shehzad Bashir, Zoe Blumenfeld, Scott Booth, Henry Brill, Yena Cho, Henry de Valence, Adam Farina, Johannes Hermle, Zoe Hitzig, Hasnain Lakhani, Eva Lyubich, Kerry Persen, Kristina Pham, Sarah Pollack, Hannah Pritchett, Georg Rilinger, Santi Ruiz, Szymon Sacher, Florian Scholz, Meg Tong, Monika Tuchowska, Helen Wang, Kim Withee, Charles Yang, and Jack Clark.
Works cited
Acemoglu, Daron, and Pascual Restrepo. 2019. “Automation and New Tasks: How Technology Displaces and Reinstates Labor.” Journal of Economic Perspectives 33 (2): 3–30. https://doi.org/10.1257/jep.33.2.3
Beraja, Martin, and Nathan Zorzi. 2025. “Inefficient Automation.” Review of Economic Studies 92 (1): 69–96. https://academic.oup.com/restud/article-abstract/92/1/69/7612958
Demirer, Mert, Leon Musolff, and Liyuan Yang. 2026. “Writing Code vs. Shipping Code: Productivity Effects Across Generations of AI Coding Tools.” NBER Working Paper 35275. https://doi.org/10.3386/w35275
Ewens, Michael, and Xavier Giroud. 2025. “Corporate Hierarchy.” NBER Working Paper 34162. https://doi.org/10.3386/w34162
Favaro, Marina, and Jack Clark. 2026. “When AI Builds Itself.” The Anthropic Institute. https://www.anthropic.com/institute/recursive-self-improvement
Hitzig, Zoe, Maxim Massenkoff, Eva Lyubich, Shaoyi Zhang, Ryan Heller, and Peter McCrory. 2026. “Agentic Coding and Persistent Returns to Expertise.” Anthropic, June 16. https://www.anthropic.com/research/claude-code-expertise
Hosseini Maasoum, Seyed Mahdi, and Guy Lichtinger. 2025. “Generative AI as Seniority-Biased Technological Change: Evidence from U.S. Résumé and Job Posting Data.” SSRN Working Paper 5425555. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5425555
Huang, Saffron, Bryan Seethor, Esin Durmus, Kunal Handa, Miles McCain, Michael Stern, and Deep Ganguli. 2025. “How AI Is Transforming Work at Anthropic.” Anthropic, December 2. https://www.anthropic.com/research/how-ai-is-transforming-work-at-anthropic
Kim, Hyunjin, and Rembrand Koning. 2026. “AI-Native Firms.” Harvard Business School Working Paper 26-090. https://www.hbs.edu/ris/Publication%20Files/26-090_96f92aa0-37d9-4789-beaa-5c0cb87a4032.pdf
Roodman, David, and Maxim Massenkoff. 2026. “An Evidence Review of Subsidized Job Training.” The Anthropic Institute Working Paper 2026-01. https://www.anthropic.com/research/reviewing-the-evidence-on-worker-retraining-programs
