PaperlayerView exact revision on arXiv

Source-linked AI summary

The Shift to Agentic AI: Evidence from Codex

Drew Johnston, David Holtz, Alex Martin Richmond, Christopher Ong, Prasanna Tambe, Aaron Chatterji

arXiv:2606.26959v1econ.GNq-fin.EC

TL;DR

Agentic AI shifts AI use from conversational assistance toward delegating work, raising questions about how people organize tasks around systems that act on their behalf. The paper analyzes Codex usage across individual, organizational, and OpenAI users, finding that intensive use increasingly involves delegation, supervision, and workflow coordination.

  • Problem

    Conversational measures such as chats and message volume may not capture agentic use, where users delegate work and systems inspect context, use tools, and modify artifacts.

    Source: §1 · p. 3, §6 · p. 22

  • Method

    The paper analyzes Codex usage with automated classifiers to produce aggregated, anonymized insights across individual users, organizational users, and OpenAI workers.

    Source: §1 · p. 2

  • Results

    Intensive Codex users increasingly delegate tasks, run work in parallel, reuse codified workflows, and shift their effort toward supervision and integration.

    Source: §6 · p. 22

  • Takeaways & Limitations

    Understanding agentic AI may require tracking delegated-task complexity, runtime, workflow reuse, concurrency, and production output, not only active users, chats, or message volume.

    Source: §6 · p. 22

    OpenAI usage is not representative of a typical organization because usage is unrestricted and internal knowledge sharing and adoption conditions are unusually favorable.

    Source: §1 · p. 2

Abstract

from arXiv · show

We analyze usage data from OpenAI's Codex tool to present large-scale evidence of how agentic AI technology, which can take actions on a user's behalf, changes how people work. We use an automated, privacy-protecting pipeline to contrast usage across three populations: external personal-account users, external organizational-account users, and workers within OpenAI. We find that agentic AI usage is growing rapidly: the number of active users has grown more than fivefold in the first half of 2026, with the most rapid increase occurring outside the initial audience of software developers. Uptake is uneven: within OpenAI, Codex usage is nearly universal and has largely replaced business usage of ChatGPT. We document a similar shift to agentic tooling outside OpenAI, particularly within organizations, although external adoption remains lower and more uneven. In addition to headline usage figures, we observe measures of sophistication, and find that a growing number of users have used Codex to change their workflows substantially. More than 10% of users manage three or more concurrent Codex agents at some point each week and that 26.6% use skills, which allow users to share instructions for complex workflows. Alongside these changes in usage practices, request complexity has increased: since the start of the year, the share of individual Codex users who submit at least one request for a task estimated to require more than eight hours for an experienced human to complete has increased nearly tenfold. Concurrently, output has grown rapidly -- in June 2026, the median OpenAI employee in a legal role generated 13 times more monthly output tokens across Codex and ChatGPT than they did in November 2025, while the median researcher generated more than 50 times as many. We conclude by discussing the implications of these patterns for productivity, job reorganization, and workforce restructuring.

1 Introduction

The paper studies how agentic AI changes work by analyzing Codex usage across individual users, organizational users, and OpenAI workers. It documents a shift from conversational assistance toward delegated production, including broader tasks and more organized workflows.

1 Introduction

  • Codex usage is analyzed across individual users, organizational users, and OpenAI workers, whose unusually favorable adoption conditions make them a frontier rather than representative case.The analysis uses automated classifiers to produce aggregated, anonymized insights without researchers reading underlying messages.Source: §1 · p. 2
  • The shift to agentic AI is rapid but uneven, with weekly active Codex usage increasing sharply over the first six months of 2026 while remaining less broadly used than ChatGPT.The shift is smallest among individual users, larger among organizational users, and largest among OpenAI workers.Source: §1 · p. 2
  • 99.8% of OpenAI workers’ output tokens across Codex and ChatGPT came from Codex as of June 11, 2026, compared with 63.3% for organizational users and 16.5% for individual users.Codex largely replaced ChatGPT as OpenAI workers’ main interface for work-related AI use.Source: §1 · p. 3
  • Codex use is oriented toward delegated production: users ask it to perform concrete work, and the share delegating tasks estimated to require more than a full day of human work has increased over time.Examples include debugging, refactoring, validation, application configuration, document drafting, and data analysis.Source: §1 · p. 3
  • Codex use remains anchored in software production but extends across its lifecycle and, especially among OpenAI workers, into research, planning, communication, data analysis, and other non-engineering work.Intensive users are more likely to use skills, run longer and more complex tasks, and coordinate multiple agents as part of repeatable workflows.Source: §1 · p. 3

2 Related literature

The paper extends research on generative AI adoption by examining agentic tools through delegated workflows and actions, rather than conversations alone. It relates these patterns to organizational complements that shape diffusion and the value of delegation.

2 Related literature

  • The paper extends generative-AI adoption research by treating a delegated workflow, rather than a conversation, as the relevant unit of analysis for agentic tools.Codex interactions make it possible to observe actions performed on a user’s behalf.Source: §2 · p. 4
  • Studies of agentic products examine broad task use and show that more autonomous tools can perform more work, reduce completion time, and expand the tasks users attempt (Yang et al. 2025; Yang et al. 2026).Coding-agent studies document heterogeneous use, user correction, and differences from benchmark-style evaluations (Baumann et al. 2026).Source: §2 · pp. 4–5
  • Research on technology diffusion emphasizes that productivity gains depend on complementary investments in business processes, worker skills, organizational design, and intangible capital.Work on agentic systems highlights supervision, verification, coordination, domain expertise, and workflow structure as relevant to value creation (Hitzig et al. 2026; Tambe 2026; Demirer et al. 2026a, 2026b).Source: §2 · p. 5
  • Workplace evidence finds that AI’s labor-market effects and task changes vary with workflow redesign and whether jobs involve automation or human–AI collaboration (Dillon et al. 2025).The paper connects these organizational-complement mechanisms to differences across individual, organizational, and OpenAI users.Source: §2 · pp. 5–6

3 Who Uses Agentic AI?

Codex adoption is broadest and most intensive among organizational users and OpenAI workers, while individual adoption remains limited. Across roles, growth extends beyond developers, and within organizations use appears across seniority levels.

3 Who Uses Agentic AI?

  • Weekly active Codex users increased more than fivefold between January 1 and June 1, 2026, and the paper focuses on Codex’s share of output tokens because user counts alone do not capture the tool shift.Many users continue using ChatGPT, while Codex adopters often use it more intensively than users who remain on conversational interfaces.Source: §3 · p. 6
  • 17.3% of organizational users used Codex in the preceding 28 days, compared with fewer than 1% of active individual users, while almost all active OpenAI workers used it weekly.Codex accounted for more than 99% of recent output tokens among OpenAI workers and a majority among organizational users; individual adopters used it intensively relative to their active-user share.Source: §3 · p. 6, Figure 1 · p. 7
  • Growth in Codex users is faster among non-developers than developers, though developers remain an important share, especially in individual and organizational accounts.Non-developers pool Personal and General Knowledge Worker personas; personas are based on the most common request label over the prior 30 days.Source: §3 · pp. 7–8, Figure 2 · p. 8
  • Across job functions, Codex’s output-token share is rising fastest in technical roles; organizational engineers average 26.8% of tokens on Codex, while legal users average 1.9%.Among organizational users who adopt Codex, it produces 88.3% of engineers’ total tokens and 17.6% of legal users’ total tokens, reflecting greater intensity among adopters.Source: §3 · pp. 8, 10, Figure 3 · p. 9
  • Within OpenAI, Codex use rose across departments, with later-adopting functions shifting rapidly from roughly 20% to 75% of output tokens within a month.Usage exceeded 90% for the average engineer by March 2026 and for almost all departments by the present day; the rapid convergence coincided with internal training and feedback loops.Source: §3 · p. 10
  • Codex’s output-token share rises across the organizational seniority distribution, indicating use among both junior and senior workers.The section’s conclusion is that adoption is broadest and most intensive among organizational users and OpenAI workers, spreads beyond technical roles, and spans seniority levels.Source: §3 · p. 10, Figure 4 · p. 11

4 What kinds of work do people do with Agentic AI?

The section measures delegated tasks with a two-level taxonomy and estimates task complexity for a sample of Individual users. Codex use is concentrated in software production, while task mix varies across worker groups and requests have become more complex.

4.1 Task shares for Codex usage

  • Across account categories, software-production tasks—including engineering operations, code implementation, code understanding, application management, and code validation—make up the largest task categories.Codex use spans the software-development life cycle, including understanding existing code, validating changes, managing applications, and supporting engineering workflows.Source: §4.1 · p. 12, Figure 5 · p. 13

4.2 Task complexity on Codex

  • The share of Individual users submitting a task estimated to take an experienced human at least 8 hours rose from 2.1% to 25.6%.The share submitting at least one task estimated at one hour or more increased from 35.4% in December 2025 to 70.2% in May 2026.Source: §4.2 · pp. 12–13
  • The first thread turn is more than twice as likely as the fourth to require work estimated at over one hour for an experienced human.The analysis estimates completion time for a 0.1% random sample of Individual accounts whose users opted into training use.Source: §4.2 · p. 13, Figure 6 · p. 14, §4.3 · p. 14

4.3 Task shares by worker type

  • Task mix varies by seniority without a simple junior-code-generation versus senior-management split; organizational executives and managers allocate a larger token share to knowledge artifacts.Within OpenAI, managers and executives devote larger token shares to collaboration-oriented tasks, while individual contributors and senior individual contributors allocate relatively large shares to engineering operations.Source: §4.3 · p. 14
  • Within OpenAI, engineering operations is common across developer and non-developer roles; in organizations, knowledge-artifact generation is common in sales, marketing, and recruiting.Most usage in both settings is concentrated in engineering operations.Source: §4.3 · p. 14, Figure 7 · p. 15

5 How Are People Using Agentic AI?

The section examines how Codex work is organized through concurrent agents, longer active runtimes, and reusable skills and plugins. These practices are much more pronounced among OpenAI users than external users, while skills are increasingly used across account categories.

5.1 Turn Concurrency

  • Nearly 28.6% of OpenAI users managed five or more concurrent agents during the week before June 11, 2026, compared with 67.4% of Organizational and 63.9% of Individual users who used no concurrent turns.Only 10.7% of OpenAI users used a sole workflow at any one time during that period.Source: Figure 9 · p. 17, §5.1 · p. 18

5 How Are People Using Agentic AI?

  • Concurrency, runtime, and systematization are observable signs of movement beyond one-off assistance, but they do not fully capture organizational redesign.The paper notes that workers and organizations can experiment with new workflows at relatively low cost, while full organizational complements are still emerging.Source: §5 · p. 17

5.2 Long-running agents

  • On June 11, 2026, the median OpenAI employee had Codex turns active for 2.5 hours, while the 99th-percentile user accumulated about 71 hours of agent turns in an average day.Since April 7, 2026, 99th-percentile users’ cumulative daily runtime increased nearly 88%; at the 99th percentile, runtime also rose roughly 25% among Organizational users and about 50% among Individual users.Source: §5.2 · p. 18, Figure 10 · p. 19

5.3 Systematization of Agentic Work

  • Skill use rose from 5.4% of active Codex users on March 1, 2026 to 26.6% on June 11; it reached 96.2% among active OpenAI users, 30.4% among Organizational users, and 25.7% among Individual users.The account-category figures cover the 7-day window ending June 11, 2026.Source: Figure 11 · p. 20, §5.3 · p. 21
  • Plugins extend Codex into recurring task domains, while custom skills encode local procedures and preferences for repeated tasks too specialized for standardized plugins.Plugin and custom-skill use is higher among Organizational users than Individual users, consistent with greater relevance of organization-specific context and shared conventions in organizational settings.Source: §5.3 · pp. 21–22

6 Conclusion

Agentic AI is shifting workplace AI use from conversation toward delegated production and new work practices. In OpenAI’s low-friction setting, use has spread across functions alongside sharply higher output, though the evidence may not represent typical organizations.

6 Conclusion

  • Agentic systems shift AI use toward delegated work: producing artifacts, modifying systems, and executing workflows rather than mainly exchanging information.The paper suggests standard measures such as active users, chats, or message volume may become less informative, and identifies task complexity, runtime, workflow reuse, concurrency, and production output as measures for future analyses.Source: §6 · pp. 22, 24
  • Intensive users assign tasks, review outputs, coordinate parallel work, and reuse codified workflows, shifting their effort toward supervision and integration.The paper characterizes adoption as the development of new work practices, not simply more frequent AI use.Source: §6 · pp. 22, 24
  • Software remains the leading use case, but Codex is also used for debugging, validation, documentation, configuration, and operational work.The paper describes software work as digital, verifiable, and modular, properties that may lend themselves to agentic AI.Source: §6 · p. 22
  • As adoption deepens, agentic work expands across research, planning, communication, recruiting, sales, product work, and data analysis at OpenAI.The paper cautions that OpenAI’s evidence may not represent typical organizations, while describing its setting as one with low adoption frictions.Source: §6 · pp. 22–23
  • At least 10-fold, median worker output tokens rose in every OpenAI job function from November 1, 2025 to June 11, 2026.The measure includes ChatGPT and Codex output tokens; the paper attributes growth to several factors, including the shift to agentic tools and more intensive, persistent use.Source: §6 · p. 23, Figure 12 · p. 23
  • Organizations may need to reconsider work allocation, review, and coordination as AI use shifts from consultation to delegation.The paper suggests jobs may place more emphasis on judgment, oversight, coordination, and review, and that changes could affect team composition, hiring, career ladders, and skill-level work distribution.Source: §6 · p. 23

A Additional Figures

The appendix figures provide additional views of Codex use by manager status, account category, persona, and task area. Their accompanying descriptions specify the measures, samples, and classification rules used to interpret those comparisons.

A Additional Figures

  • Figure A1 compares Codex usage share by people-manager status; the supplied description defines the measure as each series’ average user-level share of output tokens across ChatGPT and Codex among organizational users with job-title coverage.It uses output from the preceding 28 days and includes users active in that period.Source: Figure A1 · p. 26, A Additional Figures · p. 26
  • Figures A2–A4 show first- and second-level task-category use by account category, distinguishing users’ category reach from their average task mix.A2 and A3 count users with at least one classified turn in a category; A4 reports the average user-level share of turns by second-level category. The descriptions specify a 4% user sample, exclude “other” turns, and report 95% cluster-bootstrap intervals.Source: Figure A2 · p. 27, A Additional Figures · pp. 27–29, Figure A3 · p. 28
  • Figure A5 compares Codex use cases by people-manager status for OpenAI workers and organizational users with job-title coverage.The supplied text identifies the populations and manager-status comparison but does not describe specific outcomes.Source: Figure A5 · p. 30, A Additional Figures · p. 30
  • Figures A6–A7 show task shares and persona mix, with persona labels assigned using users’ most frequent recent request-level labels.The described persona chart uses the first and second labels, shows shares within account-category columns, excludes OpenAI users, and omits personas below 0.3% in the relevant category.Source: Figure A6 · p. 31, A Additional Figures · pp. 31–32, Figure A7 · p. 32
  • Figure A8 presents skill use by task area; its description defines each dot as the share of classified conversations invoking at least one skill during the seven-day window ending June 11, 2026.It distinguishes OpenAI, organizational, and individual accounts and reports 95% user-cluster bootstrap intervals.Source: Figure A8 · p. 33, A Additional Figures · p. 33

B Additional task-usage measures and second-level tasks

The appendix complements average task mixes with measures of how widely categories are used, then examines more granular activities. These views show Codex spanning technical workflows and knowledge production, with task emphases differing between OpenAI and organizational users.

B Additional task-usage measures and second-level tasks

  • The share of users touching a task category at least once measures breadth, complementing the average user-level task mix, which gives each active user equal weight.A category can reach many users while accounting for little of their activity, or reach fewer users while occupying a larger share of their activity.Source: B Additional task-usage measures and… · p. 34
  • Figures A3 and A4 extend the task analysis to a second-level taxonomy, reporting category reach and the average user’s distribution of work across more specific activities.These measures distinguish how many users touch a specific task from how the typical user allocates classified turns across such tasks.Source: B Additional task-usage measures and… · p. 34
  • Second-level task categories show Codex used across granular activities including debugging, environment setup, documentation, planning, repository management, data analysis, and research.The paper characterizes this breadth as movement across stages of software and knowledge production, from understanding problems through configuration, implementation, and validation.Source: B Additional task-usage measures and… · p. 34
  • Common second-level activities include code Q&A, environment configuration, documentation, repository management, code configuration, and bug fixing, extending use beyond writing new code.The paper interprets their prominence as consistent with Codex being used as a general technical assistant for surrounding development and maintenance tasks.Source: B Additional task-usage measures and… · p. 34
  • OpenAI users are more represented in web search, research, internal knowledge, application management, and some operational activities, while organizational users skew toward direct software-development tasks.The comparison describes organizational users as relatively more represented in code Q&A, implementation, repository management, code configuration, and environment configuration.Source: B Additional task-usage measures and… · pp. 34–35

C Job Title Classification

The job-title classifier uses supplied titles to infer department, seniority, people-manager status, and a cleaned job-title class.

C Job Title Classification

  • The job-title classifier uses only the supplied title and returns inferred department, seniority, people-manager signal, and cleaned job-title class.The prompt includes definitions for each returned field.Source: C Job Title Classification · p. 35

C.1 Job Title Classification Prompt

The prompt classifies organizational users’ job titles along department, seniority, people-management, and cleaned occupation dimensions. It defines distinct output labels and maps fine-grained results into coarser display categories.

C.1 Job Title Classification Prompt

  • The classifier separates inferred department, role level, people-manager signal, and cleaned job-title class into distinct outputs.Department denotes likely work area rather than seniority or literal title family; manager status is separate because a Manager title does not always imply managing people.Source: §C.1 · p. 35
  • Department labels cover functions such as Engineering, Legal, Research, Finance, Healthcare, and Other / Unknown.The prompt supplies guidelines for assigning functions and reserves Other / Unknown for seniority-only, malformed, vague, or otherwise uninformative titles.Source: §C.1 · pp. 35–37
  • Seniority categories range from Student / Trainee / Intern through IC and manager ranks to executive and founder levels.The prompt also defines people-manager labels and a separate cleaned occupation class, including generic occupations such as Consultant and Analyst.Source: §C.1 · pp. 35, 37
  • Before plotting, fine-grained department, title-class, seniority, and manager-status outputs are collapsed into coarser display categories.For example, Manager / Team Lead and Director / Head become Manager / director, while ambiguous manager titles remain Unknown / ambiguous.Source: §C.1 · pp. 38–39

C.2 Job Title Classifier Validation

The appendix reports common observed titles within classifier categories as a validation check, with disclosure-based redaction. It then presents a persona taxonomy spanning developer, general knowledge worker, and personal use labels.

C.2 Job Title Classifier Validation

  • The validation lists each category’s most common titles but redacts titles that fail a disclosure rule based on occurrence across organizations.A title is displayed only if observed in at least two distinct non-OpenAI organizations or in OpenAI’s internal sample; Unknown and Placeholder / Malformed titles are omitted.Source: §C.2 · p. 39
  • Observed titles align with diverse classifier categories, including Consultant / professional services, Engineering / technical practitioner, and Research / scientist role.Examples include Consultant at 29.6% within Consultant / professional services, Senior software engineer at 7.7% within Engineering / technical practitioner, and Quantitative researcher at 3.1% within Research / scientist role.Source: §C.2 · pp. 39, 40, 42

D Persona Prompt

  • The persona prompt’s taxonomy distinguishes developer specialties from general knowledge-worker functions and personal-use categories.Labels include developer | frontend_engineer, general_knowledge_worker | legal_policy, and personal | household_admin.Source: D Persona Prompt · pp. 45–46

E Task Classifier Prompt

The task-classification prompt organizes Codex requests into task types spanning software work, knowledge artifacts, collaboration, business functions, and application management. Its labels distinguish technical activities from broader workplace workflows.

E Task Classifier Prompt

  • The taxonomy covers code implementation, validation, understanding, and engineering operations, including building, testing, repository changes, and production investigation.It also distinguishes implementation tasks such as bug fixing from operational work such as build and release pipelines and environment configuration.Source: E Task Classifier Prompt · pp. 46–47
  • Data analysis and research labels include queries, interpretation, processing, labeling, internal knowledge, market research, financial markets, and web search.The prompt also classifies deliverables such as documents, presentations, spreadsheets, multimedia assets, and PDFs as knowledge artifacts.Source: E Task Classifier Prompt · pp. 47–48
  • Collaboration categories cover messages, summaries, scheduling, task management, and meeting support.Business-function workflows separately cover finance, sales, marketing, product and design, customer support, recruiting, and legal and compliance.Source: E Task Classifier Prompt · pp. 48–49
  • Application-management labels include configuring Codex, data entry, and browser actions performed on a user’s behalf.Codex configuration includes instructions, skills, apps, connectors, MCP servers, automations, and model settings.Source: E Task Classifier Prompt · p. 49

F Query Complexity Prompt

The paper estimates how long an experienced human would need to complete the substantive request in a Codex conversation without AI, counting hands-on work rather than waiting or runtime. A Codeforces benchmark finds the estimates correlate with two metadata-based difficulty measures, while tending to exceed them.

F Query Complexity Prompt

  • The complexity classifier estimates active human work time, in minutes, to complete the substantive request without AI.It excludes waiting, approvals, builds, and AI runtime; when a transcript contains only wrapper text, it estimates the embedded request, and returns zero when no substantive request is present.Source: F Query Complexity Prompt · p. 49
  • Correlations are 0.65 between model-predicted time and the five fastest human solvers’ average completion time, and 0.69 between Codeforces difficulty score and model-predicted time.The benchmark uses 1,000 random Codeforces problems; the classifier generally predicts slower completion times than metadata measures, which the authors partly attribute to comparison with a selected group of competitive programmers.Source: F Query Complexity Prompt · pp. 49–50, Figure A9 · p. 50

Related papers

Loading 2606.26959v1…