amplifying/research · jun-2026
State of the Coding Agent Market
June 2026 edition: 88.3M units of attributed public output, and who is growing, who gets merged, and who reviews the work, measured from primary data.
Every figure here is live on the Coding Agents Index, updated daily.
What the data says
- 1.Visible agent output is top-heavy and led by Claude Code. Of 88.3M attributed commits and pull requests on public GitHub, the top three (Claude Code, Replit Agent, and Cursor) account for 87%, and Claude Code alone is 71%. Six agents clear one million units; the other 16 attributable agents split the remaining 2%. Units differ by agent (Codex is counted in pull requests, Replit in its per-prompt checkpoint commits), so read this as a concentration signal, not a like-for-like ranking.
- 2.Agents already mark a measurable share of all public PR flow. Sampling the 7998 most recent public PRs on GitHub, 17.4% carry a public agent signal, exact counts rather than estimates, and Claude Code alone marks 13.2% of everything sampled.
- 3.Enterprise readiness is now table stakes at the top. Of 13 agents with published enterprise profiles, 11 hold SOC 2 and 11 ship SSO/SCIM. Self-hosting remains rare: 3 options.
- 4.Most merged agent work still passes a human, but it splits sharply by agent. Across 230 merged PRs in significant repos, 55.7% had a human reviewer. Cursor and the Copilot agent route nearly everything past review; Codex and Devin self-merge the majority. Trust, not volume, is where the agents actually diverge.
Market structure
The cleanest public measure of coding-agent adoption is output you can count: commits signed by the agent, commits visible through the agent's own bot or workflow, or pull requests carrying hosted tasks. As of 2026-06-09, the index attributes 88.3M such units across 22 agents.
Attributed public output, snapshot 2026-06-09
Counts are floors. Units differ where noted: Codex is measured in cloud pull requests because it does not sign commits; each PR carries at least one commit. Methodology.
The market is concentrated: the top three agents hold 87% of visible output, the six above one million units hold 98%, and Claude Code alone is 71%. That reflects both adoption and measurement: terminal and hosted agents leave public signatures by default, while IDE-first work often lands under the developer's own identity.
Attributed output rolled up to the owning company
Blue bars are listed companies; green are private. Live view on By Owner.
Rolled up to the companies that own the agents, 19% of all attributed output sits inside listed companies (Microsoft, Alphabet, Amazon); the rest belongs to private companies, led by Anthropic. For public-market investors, most of the visible agent economy is not yet buyable.
Trends: weekly agent-marked PRs
Weekly counts of agent-marked pull requests created on public GitHub, from date-windowed search. Agents visible through their own bot or workflow are exact; agents measured through declared attribution are estimates. Defaults to linear; switch to log when lines span several orders of magnitude. Window: 2025-02-03 to 2026-06-08.
| Agent | PRs, latest week | Move | Commits, latest week |
|---|---|---|---|
| Claude Code | 455K | +48% / 4w | 3.7M |
| Copilot agentexact | 16,869 | -56% / 4w | 24,914 |
| Cursor agents | 28,462 | +31% / 4w | 354K |
| Devinexact | 8,291 | -13% / 4w | 4,950 |
| Julesexact | 56,720 | +10% / 4w | 17,276 |
| Codex | 125K | -18% / 4w | n/a |
The shape of the market in one chart: Claude Code went from 18,800 marked PRs a week in December 2025 to 455K in the latest complete week, about a 20x climb in six months. Jules is the quiet riser, 21x over nine months. Codex holds a high plateau in the 135,000 to 155,000 band. The Copilot coding agent is the one measured decline that is not a measurement artifact: its counts are exact bot data, and they have halved since December. Trend data updates daily via the capture pipeline.
The six that matter
Weekly marked PRs, latest complete week (2026-06-08), with each player's trajectory and the measurement caveat that matters for reading it.
Anthropic (Claude Code), 455K PRs/week. The market leader by an order of magnitude, and still compounding: 7,700 a week last September, 18,800 in December, 455K now. In the live flow sample it marks 13.2% of the 7998 most recent public PRs, more than dependency bots like dependabot and renovate combined, and more than every other agent put together. The caveat runs the other way for once: squash merges strip its markers, so this is the floor.
OpenAI (Codex), 125K PRs/week. The clear number two, with 5.1M PRs all-time and the highest merge rate of the big cloud agents at 86%. But the line has been flat in a 135K-155K band since spring while Claude Code tripled: Codex held the weekly lead as recently as last September (82K to Claude's 7.7K) and has since been passed roughly 2.7-to-1. Its CLI work leaves no public mark, so the subscription-driven CLI business is invisible here.
Google (Jules), 56,720 PRs/week, the surprise of this edition. Jules grew 21x in nine months: 2,600 weekly PRs in September 2025, 13,700 in December, 56,720 now, passing the Copilot coding agent during March. Monthly output went 12,900 (September) to 230,000 (May). What makes Jules the anomaly is that one profile holds across every cut. It is the most globally distributed agent we track, with only about 11% of its located users in the US (India is also 10%); it leads the Flutter framework tilt (24.4%) and the Education industry tilt (34.3%); and it has the lowest org-repo share of the six at 10%. Read together, that is a free-tier, global, consumer-and-hobbyist footprint at scale, not enterprise penetration. Google's other product, Gemini CLI, barely registers (about 17,000 signed commits a month, flat) because it does not sign work by default; Google's real footprint is likely much larger than its mark.
GitHub (Copilot agent), 16,869 PRs/week and falling. The only major in measured decline: 45K in December, 43K in early May, 16,869 in the latest week, on exact bot/workflow counts, so the halving is real, not an attribution change. The shape of its work is the most enterprise-like in the index: it leads the .NET tilt at 40.5% and Gaming at 33.9%, and its users skew toward Germany and the UK. The structural caveat is also the strongest here: Copilot is sold into private enterprise repos that public data cannot see, and its 4.7M paid seats say the business and the public mark have decoupled.
Cursor, 28,462 background-agent PRs/week. The most under-measured major: only its background agents leave PR marks, while its editor-led volume, the bulk of its roughly $4B run-rate, committed invisibly until 2026. That changed with its Agent Trace attribution push: Cursor signed almost nothing before February 2026, attributed roughly a million commits a month through spring under one format, and now signs 250K-350K commits a week under another, so its visible series is younger than the product. Where it shows up, it concentrates: it leads React (29.2%), Django (80.5%), Crypto/Web3 (40.9%), and Devtools (24.2%) tilts, with a user base outside the US led by Germany (8.6%), China (7.5%), Austria (5.4%), Israel (5.4%).
Devin (Cognition), 8,291 PRs/week. Small next to the platforms but growing 13.7x year-on-year from 680 a week last September, with the most US-centric user base of the six (12.5% of located users) and a portfolio that reads like a services firm: it leads Angular (43.6%) and Rails (12.2%) plus AI/ML (19.3%) and Finance (25.3%). Cognition also owns Windsurf, whose editor work is unmarked, and enterprise private-repo work is invisible, so this also reads as a floor.
Autonomy and trust
Volume says how much agents ship. The harder question for a buyer is whether anyone checks the work. We sample merged agent pull requests and ask three things public GitHub can answer: did a human other than the author review it, how long did it sit before merging, and did it touch a test. To keep the read honest we restrict the review rate to significant repos, org-owned or with at least ten stars, because solo self-merge in throwaway personal repos otherwise swamps everything. Across 230 such PRs (2026-05-07 to 2026-06-05), 55.7% had a human reviewer and 44.3% merged with none.
Human-reviewed share of merged PRs, significant repos
Rates capped per repo before computing. 1 agents are withheld here: their samples are too concentrated in a few auto-merge repos to publish a fair rate. Live view on Autonomy.
The agents diverge sharply here. Cursor and the Copilot agent route almost everything past a human; their merged work behaves like a developer's own branch awaiting approval. Codex and Devin sit near the other end, merging the majority of their PRs with no second reviewer. Two agents can ship comparable volume and still differ this much in how supervised that work is.
Time-to-merge confirms it rather than contradicting it. Reviewed PRs sit longer by design: Devin's reviewed PRs take a median 2.15h versus 14 min when unreviewed, so a low median merge time is usually a sign of auto-merge, not speed. Test-touch is the quieter signal: Claude Code 42.5%, OpenAI Codex 40.8%, Cursor 35.8%, GitHub Copilot 28.3% of merged PRs touch a test file. The agents whose work clears review are also likelier to ship the tests with it.
Merging is not the same as reviewing. Settled merge rates run high across the board: Claude Code 91.5%, Codex 87.1%, Copilot 80.5% of recent PRs, but a high merge rate paired with a low review rate means the agent is trusted to self-merge, not that a human signed off. For a buyer evaluating where agents can run unsupervised, the review column is the one that matters.
What agents build
Across the 32 tracked languages, 18% of all public PRs in the 2026-05-01 to 2026-05-31 window carried an agent mark. The spread tells you where adoption actually lives. Swift is the most agentified language at 30.7% of all its PRs. The laggards are where the next wave of adoption has to come from: Ruby sits at just 5.6% despite 155K monthly PRs.
Month over month (2026-05 vs 2026-04), the fastest-agentifying stacks are Elixir (+81%, 4,876 PRs/mo), R (+69%, 4,139 PRs/mo), Solidity (+54%, 2,561 PRs/mo). A live read on where adoption is accelerating right now.
By raw volume Claude Code leads every single language, from 37.9% of agent-marked JavaScript to 73.9% of agent-marked Shell. Codex is its only volume rival, strongest in Rust, C++, and HTML. The sample-based framework tilts show where the others specialize: Cursor leads Django (80.5%) and React (29.2%); the Copilot agent leads .NET (40.5%); Jules leads Flutter (24.4%); Devin leads Angular (43.6%) and Rails (12.2%). React and Next.js together draw more sampled agent work than the next ten frameworks combined.
The industry cut has one headline: agents overwhelmingly build AI. 42% of classified sampled volume lands in AI/ML repos, with Devtools (12%) and Finance (6%) far behind. The agent economy is, so far, largely feeding itself. The niche leaders are telling: Devin leads AI/ML and Finance, Cursor leads Crypto/Web3 and Devtools, the Copilot agent leads Gaming, and Jules leads Education, consistent with its free-tier, global, non-enterprise profile.
Where the users are
Country shares below reflect the June 2026 edition snapshot; the live map currently covers 2026-06-13 to 2026-06-19.
From daily samples of PR authors with public profile locations (a minority of users, so directional): the US holds 9.2% of located agent users overall, but the six majors have sharply different geographies. Devin is the most American at 12.5% of its located users. Codex skews to Japan (9.4%) and China (9.4%). The Copilot agent skews European, with Germany (11.6%) and the UK (10.6%) as its second and third markets. Cursor's largest non-US pockets are Germany (8.6%), China (7.5%), Austria (5.4%), Israel (5.4%). Claude Code's most distinctive market is South Korea, its number two country at 4.2%. Jules is the least US-dependent agent we track at about 11%, tied with India, which fits a free product riding Google's global distribution.
Read together with the org-repo data, a pattern emerges: the agents with enterprise business models (Copilot, Devin, Cursor) concentrate in the US and Europe, while the free-tier agents (Jules, and Gemini CLI where visible) are growing fastest in markets the paid agents have not converted. The geographic frontier of this market is priced in requests per day, not dollars per seat.
The attribution lens
Divide tracked revenue by attributable output and the market splits into two business models. The figures below are run-rate dollars per unit of visible output, a measure of how much of each company's work is countable, not how efficient it is.
Run-rate revenue per unit of attributed output
Per-unit value is derived from sourced revenue divided by sourced attributed output. Replit revenue is a stated target.
Devin sits at one end (~$1,919 per unit): it charges enterprise contracts against a bot that pushes comparatively rarely. At the other end, Claude Code (~$40) and Replit (~$121) sign by default, so far more of their work is visible per dollar. Cursor (~$698) sits in between: most Cursor usage is IDE editing that commits as the human, so only its background-agent signed commits and its 614K agent PRs surface here. The spread is a visibility gap, not an efficiency one.
Commit counts also carry a granularity bias, because agents commit at different rates per task. For the agents that open pull requests themselves, PRs are the apples-to-apples unit: one PR is roughly one task.
| Agent | Agent-opened PRs | Attributed commits | Commits / PR |
|---|---|---|---|
| OpenAI Codex | 5.1M | does not sign | n/a |
| Claude Code | 4.2M | 62.9M | 15.1 |
| GitHub Copilot | 1.9M | 3.8M | 2.0 |
| Jules | 1.2M | 1M | 0.9 |
| Cursor | 614K | 5.7M | 9.3 |
| Devin | 161K | 256K | 1.6 |
| Amazon Q Developer | 2,218 | 2,218 PRs | 1.0 |
By tasks shipped, the order is Claude Code, Codex, Copilot, Cursor, Jules, Devin. Two reads worth holding onto: Replit's 8.3M checkpoint commits disappear from this lens entirely, because Replit Agent checkpoints commits but never opens the PR; and the Claude figure is the softest in the table, a PR-template text match rather than its own workflow. Google now shows up twice, with Jules at 1.2M agent-opened PRs and a 202K-commit Gemini attribution floor that spans Gemini CLI and its other surfaces.
Breadth differs too. Among recently sampled PRs, Jules averages 9 PRs per distinct author; Devin averages 4.6 PRs per distinct author; Cursor averages 4.5 PRs per distinct author; Codex averages 3 PRs per distinct author; Copilot agent averages 2.6 PRs per distinct author; Claude Code averages 1.9 PRs per distinct author. High ratios mean a smaller base using the agent heavily; low ratios mean broad, shallower adoption.
Accumulated totals are stock. Flow is sharper: sampling the 7998 most recent public PRs on GitHub (8 windows, latest 2026-06-17), 17.4% carry a public agent signal, and Claude Code alone marks 13.2% of all sampled public PRs. Dependency bots add another 13.8%. Codex deserves a measurement note. Its count is an estimate of the pull requests its cloud service opens: well-covered for that cloud path (5.1M all time, about 86% merged), but inferred from public signals rather than an exact bot identity, which still puts it near the top of the PR-stock table alongside Claude Code. An earlier signal was retired in November 2025 and briefly made it look absent until a replacement recovered the series. Its local command-line work runs under the developer's own identity, so that output is largely invisible to public attribution and is not in this count. Lesson for reading any agent metric: signals break and recover, and the methodology page records each change.
The lens also shows who is absent. GitHub Copilot has 4.7M paid users and 3.8M attributed commits, but Microsoft does not break out Copilot revenue, so it cannot be placed on this chart. Windsurf books $82M with no attributable output at all. And Jules has 1.0M commits with no disclosed business model around them yet. Where output is high and revenue is undisclosed, either monetization is coming or the output is subsidized distribution. Both are worth watching.
The CLI shift
The index tracks 24.8M weekly npm downloads of agent CLIs as a demand floor. The caveat matters more than the number: the leading CLIs have moved to native installers (Claude Code now recommends native install, Homebrew, WinGet, apt, dnf, or apk), so npm captures a shrinking share of real installs. Channel-adjusted figures on the index use the current npm channel as a floor, with native installs unobserved.
The business significance of the CLI is composability. A terminal agent runs headless in CI, signs its commits, and stacks into pipelines, which is exactly why CLI-led agents dominate the attributed-output table: signing is natural where automation lives. The IDE agents that win on revenue per seat are invisible in the same table. Distribution model determines measurement model.
Enterprise readiness
Of 13 agents with published enterprise profiles, 11 hold SOC 2, 11 ship SSO/SCIM, and 7 offer zero-data-retention at least on enterprise tiers. Only 3 offer self-hosted or VPC deployment. The compliance baseline has commoditized at the top of the market; deployment control has not. For regulated buyers, the self-host shortlist is the actual market.
Weekly agent PRs weighted by org-repo share
The best public proxy for enterprise distribution is volume-weighted agent share inside organization-owned repos, company and team accounts rather than personal ones. On that measure Claude Code takes 73% of weekly org-repo agent PRs. Public repos only, so agents sold mainly into private enterprise repos (Copilot, Devin) are understated. The living view is on Industries.
Methodology and sources
Attributed output comes from each agent's public signal: its own bot or workflow, attribution it declares on its work, or hosted task metadata, snapshot 2026-06-09; the visibility classes, validation, and caveats are published on the index methodology page. The attribution approach matches the AgentPack dataset (arXiv:2509.21891). Revenue, funding, pricing, and enterprise fields are sourced per agent on the Coding Agents Index, each with its own source links and confidence labels. All counts are floors, all snapshots are dated, and every number here can be reproduced from its source.