On April 14, Anthropic shipped Claude Code Routines, cloud agent tasks that fire on schedule, API call, or GitHub webhook. The next day it shipped a desktop rebuild: parallel session management, drag-and-drop panels, native SSH on macOS and Linux, full plugin parity with the CLI.
By timing alone this release is late. OpenAI shipped the Codex macOS app and Automations on February 2. Cursor shipped Automations on March 5. Anthropic is ten weeks behind Codex and five weeks behind Cursor. The community read of “Anthropic is playing catch-up again” showed up fast. But look at the three products closely and the catch-up framing only holds under one assumption: that all three are doing the same thing. They aren’t. They only share a name.
The fastest way to see the difference is to ask: if each of these agents were an employee you hired, what job would they be doing?
Codex Automations is a junior developer who writes shell
scripts at the desk next to yours. Every morning it hands you a
digest from git log. Once a week it runs a skill upgrade.
That’s it. It runs on your own Mac, the Codex app must stay open, the
project files must be on disk (the official docs say
so in plain language: “The app needs to be running, and the selected
project needs to be available on disk.”). Cron only, no webhook, no
separate billing — it rides on your ChatGPT subscription. It serves an
individual developer sitting in front of a laptop. What it replaces is
the handful of crontab lines you’d have written yourself.
Cursor Automations is a DevOps engineer inside a mid-size company. On push to main it runs a security review. On a PagerDuty alert it pulls Datadog logs and correlates recent commits to produce a candidate fix. On a new Linear issue it scores PR risk and routes decisions into Slack or Notion. Triggers span Slack, Linear, GitHub, PagerDuty, and custom webhooks, and the agent runs in Cursor’s own cloud sandbox. Rippling is the first named external customer. Cursor’s engineering lead Jonas Nelle put it to TechCrunch: “It’s not that humans are completely out of the picture. It’s that they aren’t always initiating. They’re called in at the right points in this conveyor belt.” This is Cursor positioning against GitHub Actions, PagerDuty, and Zapier — serving companies that already have end-to-end automation wired up.
Claude Routines is the engineer on an enterprise CI team who writes webhook handlers and API integrations. Each routine is an independent API endpoint with its own bearer token. Enterprise CI, Jenkins, an internal orchestrator, any in-house system can invoke it via standard HTTP. Independent quotas (5/15/25 per day + overage) and independent billing let agent tasks be attributed to cost centers by department. The desktop rebuild’s SSH support means agents can run on the company’s own internal dev machines, with code never leaving the network. Routines serves companies that will fit the agent into an IT architecture diagram, run it through procurement, and put it in a signed contract.
These three jobs are not interchangeable. Codex’s cron scripts can’t do PagerDuty automation. Cursor’s switchboard is not a substitute for an addressable, programmable API primitive. Three companies, three lanes, not the same race.
Why did the three end up with three different products? Because each one is actually serving its dominant paying customer.
About 80% of Anthropic’s revenue comes from enterprise and API, with Claude Pro/Max at 15–20%. After the February 2026 Series G, Dario’s public framing was “80% of revenue from business customers.” There are 300,000+ enterprise customers, roughly 1,000 paying over $1M a year, and Claude Code is at $2.5B ARR with more than half of that from enterprise subscriptions (from Anthropic’s own Series G announcement). Consumer Claude paid subs doubled in Q1 2026, but Claude Code also doubled in six weeks, so the consumer share actually compressed slightly rather than expanding. For comparison, Claude.ai’s weekly actives are 19M against ChatGPT’s 900M — about 4.5% share, with no meaningful catch-up in eighteen months. Making Routines a contract-addressable API primitive is the natural output of serving the customer who actually pays.
70% of OpenAI’s revenue comes from ChatGPT subscriptions, with API and enterprise at ~15% each. Nine hundred million weekly actives is its biggest asset. Codex doesn’t have its own SKU; it’s embedded in ChatGPT Plus ($20/mo) and Pro ($200/mo) with message quotas. Codex Automations requiring the Mac app open with files on disk locks the user picture architecturally to “individual developer at a laptop.” It’s the 2C gene of ChatGPT extending into a developer product.
Cursor’s $2B ARR breaks down to roughly $1.2B enterprise and $800M from ~3.3M paying individuals. Enterprise customers include about half the Fortune 500 (Rippling, Brex, Ramp, Stripe, Discord, Samsara, Airtable, Sierra AI, NYT are all publicly disclosed). Michael Truell told The Verge that vibe coding and hobbyist users are “not Cursor’s primary audience.” To justify valuations moving toward $10B+, enterprise is the only road. That’s why Cursor 3 (codename Glass) demoted the IDE to a fallback surface, and why Truell wrote in the launch post that “we will continue to invest in the IDE until codebases are self-driving” — a line that explicitly calls the IDE transitional.
I wrote a piece in May 2025 called AI Products: the Misalignment looking at the same gene playing out on consumer apps: Claude.ai drops your task when you switch apps, its iOS client overheats on long outputs, chat history becomes “Untitled”. The conclusion then was “Claude is a B2B company, so the consumer product is a half-finished one.” Eleven months later the B2B side of that gene is fully visible: Anthropic didn’t go fix the consumer app, it poured everything into Claude Code and made it the second B2B revenue pillar on top of API wholesale.
Routines isn’t a standalone release. Over the past two months, Anthropic has done five things pointing in the same direction:
The runtime layer quietly lowered default reasoning effort (starting in February; only admitted after AMD’s Stella Laurenzo did a reverse audit on 6,000+ local session files in early April). Subscription OAuth was restricted to first-party products on March 21 via the legal compliance page, cutting off OpenCode the same day. Defense in Depth landed on April 1 (Zig attestation, FFI anti-debug, client obfuscation). Managed Agents shipped on April 8, taking custody of agent credentials, memory, and event history. Then Routines + desktop rebuild on April 14–15.
Any one of these moves could be explained as an isolated product decision. Together they read as a coherent project: Anthropic is systematically pushing Claude Code from “terminal CLI” to “procurable, auditable, budget-line-item agent runtime platform.” Runtime throttling protects margin. OAuth restriction keeps context inside the product boundary. Defense in Depth makes the client stable enough to be infrastructure. Managed Agents offers a hosted runtime. Routines is the addressable agent primitive. Each step maps directly to a question enterprise procurement will ask when buying an agent product: Is the cost controlled? Does the data stay in-product? Is the client stable? Can we host it? Can we integrate with our CI?
I’ve analyzed each of the first four in separate pieces: runtime layer regression, a Claude Code subscription is not a developer credential, Defense in Depth, and Managed Agents. Adding Routines to that sequence shows a full cadence of enterprise agent platform construction. Anthropic left Claude.ai mostly alone and routed its budget into Claude Code. That’s a gene-driven resource allocation.
If three companies are building three different products for three different users, why did all three push hard in Q1 2026? Underneath there’s a shared constraint tightening: cloud agents have finally become viable in production — but the conditions for viability differ by user.
The real value of a cloud agent hinges on whether the async verification loop can be closed. Inside an enterprise team, that loop is CI/CD: diff comes in, tests run, deploy to staging, pager wakes oncall — every step can be advanced autonomously, and the human closes the loop with a tap on a phone. Rippling’s Cursor morning agent works because the CI infrastructure already exists. Firdaus Fauzi’s LinkedIn description — “Cursor cloud agents hooked into full CI/CD feels like magic” — is the same observation. Gergely Orosz quoted the internal OpenAI Codex team: “structure the codebase to make it inevitable for the model to succeed.” That is exactly the precondition.
Without that loop, cloud agents collapse. Every.to’s Claude Code mobile experiment said it directly: “you can’t load the app on a phone and validate it works, so you can’t finish anything.” Bug reports on the OpenAI community forum about Codex cloud failing to open PRs describe the same failure mode: any broken link in the chain zeroes out the async advantage, forcing the user back to the desktop.
Map the condition back to the three products and it all fits. Cursor and Anthropic’s enterprise customers have CI/CD by default, so both make their cloud Automations/Routines a primitive for enterprise automation. OpenAI’s ChatGPT subscribers are mostly individual developers who don’t, so Codex Automations is designed to run locally on the Mac with the app open — letting the user close the verification loop with their own eyes. Anthropic’s SSH support extends the same idea into the enterprise intranet: the agent runs on the company’s own dev machine, verification goes through internal CI, and code stays inside the network. Each of the three product designs is an accurate response to its own users’ verification-loop capabilities.
The Q1 2026 wave of cloud agent releases shares a single underlying technical shift: the async verification loop is becoming closeable in more scenarios. But the three products end up so different that they are not interchangeable. The differences are revenue-structure-driven paths. Anthropic’s ~80% B2B revenue makes Claude Code an enterprise agent platform. OpenAI’s 70% ChatGPT subscription makes Codex a personal desktop assistant. Cursor’s 60% enterprise revenue pulls it from IDE toward enterprise orchestration switchboard.
“Does Claude Code have Routines?” isn’t the question that leads anywhere useful. Read the genes, not the timing.