The Agent Race Just Got More Visible
OpenAI launched workspace agents in ChatGPT this week, and the coverage has been predictably breathless. A new era of autonomous AI. The end of repetitive work. Your digital coworker has arrived.
The reality is more interesting and more useful than the headline. What OpenAI shipped is genuinely good. It also isn't new. And understanding the difference between the two matters if you're responsible for AI adoption inside your organization.
What Workspace Agents Actually Are
Workspace agents are persistent, cloud-hosted automations that run on a schedule or in response to a trigger, without anyone actively sitting in ChatGPT. A team builds an agent once, shares it across the organization, and it handles repeatable work in the background: pulling data every Friday and drafting a report, monitoring a Slack channel for customer feedback and routing it into tickets, screening inbound leads against a qualification rubric and updating the CRM.
They're an evolution of Custom GPTs, which OpenAI introduced last year but which never really took hold. The new version runs in the cloud rather than requiring an active session, supports connected apps and tools, and comes with enterprise governance built in: role-based access controls, admin approvals, audit logs, and centralized management. OpenAI is clearly positioning this as infrastructure for teams, not just a feature for power users.
For Business, Enterprise, and Edu plan customers, workspace agents are available now in research preview, free until early May before moving to credit-based pricing.
Claude Has Been Here for a While
If this sounds familiar to anyone using Claude Code or Claude's Cowork product, it should.
Anthropic shipped Routines in Claude Code two weeks ago. A Routine is a saved Claude Code configuration: a prompt, connected tools and repositories, and a set of connectors, packaged once and executed automatically. Routines run on Anthropic's cloud infrastructure, which means they keep running whether your laptop is open or not. They support three trigger types: scheduled cadence, API calls, and GitHub webhook events. And like ChatGPT workspace agents, you build them by describing what you want in plain language.
Claude's Cowork product for knowledge workers has had scheduled tasks for some time as well. The key difference there is that Cowork tasks run locally, so they require your machine to be on. Routines solved that by moving execution to the cloud.
So when OpenAI says they've launched "autonomous agents that run even when you're not in the app," Claude users reading carefully might think: yes, we've been doing that.
Where the Two Actually Differ
The functional overlap is real, but the differences matter depending on what you're trying to build.
The most significant gap today is sharing. ChatGPT workspace agents are designed to be org-wide resources. A team builds one agent and everyone in the organization can use it, governed by whatever permissions the admin sets. Claude Routines are personal. They run under your identity and aren't shared across a team. For organizations that want a centralized library of automated workflows that anyone can invoke, that's a meaningful limitation in Claude's current offering.
The enterprise governance story is also more developed on the ChatGPT side right now. OpenAI built admin controls, audit logs, and role-based access into the workspace agent launch from day one. That's not because Anthropic can't build it; it's because OpenAI is explicitly targeting enterprise IT departments with this release, and those buyers require that infrastructure before they'll approve anything.
On the other side, Claude Routines go deeper on developer triggers. The GitHub webhook integration lets a Routine fire automatically in response to repository events: a pull request opens, a merge happens, a release deploys. That's useful in ways that ChatGPT workspace agents don't yet support, and it reflects the fact that Claude Code was built with developers as the primary audience.
What This Means for Your Organization
The agent layer is becoming a standard part of how AI tools work, not a differentiating feature of one vendor over another. Within the next year, the ability to schedule an AI workflow, connect it to your business tools, and have it run autonomously will be table stakes across every major AI platform.
That's worth sitting with for a moment, because it changes what the real question is.
Most organizations we work with are still asking "which AI tool should we use?" or "how do we get our people to adopt it?" Those questions matter, but they're going to matter less as the capability gap between platforms narrows. The harder and more durable question is: which workflows in your organization are actually worth automating, who owns governing them once they're running, and how do you measure whether they're working?
Building an agent is now a conversation, not a project. Deciding which conversations to have, in what order, with what governance around them, is still hard. That's where the work is.
A Practical Starting Point
If you're on ChatGPT Enterprise or Business already, workspace agents are worth exploring now while they're free. Start with one workflow your team does every week that follows a consistent pattern: a standing report, a recurring research task, a weekly sync that requires someone to pull data from three places first. Build the agent, run it for a month, and watch what breaks.
If you're running Claude Code or using Claude's Cowork product, Routines are the equivalent capability for individual and developer use cases. The shared-agent feature set will come; Anthropic's track record on building toward enterprise capability is strong.
Either way, the infrastructure question is no longer the blocker. The judgment question is. Which is, honestly, where it should have been all along.