OpenAI just made its most direct pitch yet to IT departments and operations teams: workspace agents in ChatGPT are here, and they’re built to do the kind of multi-step, cross-tool work that used to require either a dedicated engineering team or an army of Zapier automations. Announced on April 22, 2026, these Codex-powered agents run entirely in the cloud, operate across connected tools, and are designed to scale actual business workflows — not just answer questions about them.
What Got Us Here
OpenAI has been quietly building toward this for a while. Codex, which recently crossed 4 million weekly users as its enterprise push began, started as a code-completion tool and has steadily evolved into something far more capable. Earlier this year, Codex picked up computer use, browsing, and memory — features that transformed it from a coding assistant into a genuine agent runtime.
The broader context matters too. Enterprise customers using ChatGPT — think large hotel chains, financial firms, healthcare networks — aren’t just asking AI to summarize emails anymore. They want it to act. They want it to pull data from Salesforce, cross-reference it with a Google Sheet, draft a follow-up, and log the outcome, all without a human babysitting every step. That’s exactly the gap workspace agents are designed to fill.
OpenAI also didn’t build this in isolation. The official announcement positions workspace agents as a natural extension of the ChatGPT platform for teams — particularly those already on ChatGPT Enterprise or ChatGPT Team plans. The timing aligns with intensifying competition from Microsoft Copilot, Google’s Gemini for Workspace, and Anthropic’s Claude, all of which are pushing hard into agentic territory.
What Workspace Agents Actually Do
Here’s the thing: the word “agent” gets thrown around so loosely in AI that it’s almost lost meaning. So let’s be specific about what OpenAI is shipping.
Workspace agents are pre-configured or custom AI agents that live inside ChatGPT and can be assigned ongoing, multi-step tasks. They’re powered by Codex under the hood, which means they can write and execute code, interact with APIs, read and write files, and chain together sequences of actions autonomously.
Key capabilities include:
- Cloud execution: Agents run in OpenAI’s cloud infrastructure, not on your local machine. Tasks persist even when you close the browser, which is a meaningful shift from session-based AI interactions.
- Tool integrations: Workspace agents can connect to common business tools — think CRMs, project management platforms, data warehouses, and communication apps — through defined APIs and connectors.
- Workflow automation: Teams can define multi-step workflows and hand them off to agents. The agent handles sequencing, error handling, and output delivery.
- Security controls: Admins can set permission scopes, limit what data agents can access, and audit agent activity. This is non-negotiable for enterprise rollouts.
- Collaboration-ready design: Agents can be shared across a workspace, so one team member can build an agent and deploy it for the whole department.
- Asynchronous operation: You don’t need to stay in the chat. Assign a task, walk away, and get notified when it’s done.
The Codex backbone is important here. Unlike a simple prompt-and-response loop, Codex-powered agents can actually generate and run code mid-task — so if an agent needs to parse a CSV, transform data, or call an API that doesn’t have a native connector, it can write the script to do it. That’s a genuinely different capability compared to most competing workflow tools.
How This Compares to Competitors
Microsoft’s Copilot agents have been building steam inside the Microsoft 365 suite, but they’re tightly coupled to Microsoft’s own tools. They work great if your company lives in Teams, Outlook, and SharePoint. The moment you need to reach outside that stack, things get complicated.
Google’s Gemini-powered coding and automation tools are impressive and improving fast, but they’re similarly anchored in Google Workspace. Anthropic’s Claude has strong reasoning but hasn’t shipped a comparable enterprise agent platform with this level of workflow specificity.
OpenAI’s play is tool-agnostic by design. If your company uses a mix of Slack, Notion, HubSpot, and a custom internal database — which describes most mid-size companies — workspace agents are built to reach across all of it. Whether the connector library is deep enough at launch to back that up is a real question, but the architecture is pointed in the right direction.
What This Means for Enterprise Adoption
For IT and Operations Teams
The security model is going to be the first thing IT asks about, and OpenAI appears to have thought carefully about this. Role-based access controls, audit logs, and defined permission scopes for each agent mean that enterprises can give employees agency without giving them the keys to the kingdom. That’s a baseline requirement, not a differentiator — but it’s good to see it addressed upfront.
We’ve already seen how large organizations approach this kind of rollout. Hyatt’s approach to ChatGPT Enterprise shows that the real challenge isn’t access — it’s governance, training, and trust-building across a distributed workforce. Workspace agents raise the stakes because they’re not just generating text, they’re taking actions with real consequences.
For Developers and Power Users
If you’re already building on OpenAI’s infrastructure, workspace agents open up interesting new territory. The Codex runtime means you can essentially embed custom logic inside a workflow without needing to stand up a separate service. Pair that with the Agents SDK’s native sandboxes and smarter execution and you’ve got a fairly complete stack for building internal tooling without a ton of overhead.
I wouldn’t be surprised if we see a cottage industry of pre-built workspace agent templates emerge quickly — similar to how Notion templates and Zapier zaps became their own mini-economy. The combination of a large existing ChatGPT user base and a shared agent library inside workspaces is a strong foundation for that.
For Regular Business Users
Here’s where I think the pitch is most interesting — and most uncertain. The promise is that a non-technical employee can describe what they need done, and a workspace agent handles it end to end. That’s compelling. But agentic AI still fails in frustrating ways when task descriptions are ambiguous or when edge cases arise mid-workflow. The user experience of debugging a failed agent task isn’t obvious yet.
The companies that will get real value from this early are the ones with clear, repeatable processes that currently require someone to manually stitch together data from multiple tools. Think weekly reporting pipelines, lead qualification sequences, invoice processing, or compliance documentation. Structured tasks with defined outputs — that’s where agentic AI earns its keep right now.
Key Takeaways
- Workspace agents are Codex-powered, which means they can write and execute code as part of a workflow — not just follow rigid scripts.
- They run asynchronously in the cloud, so tasks continue even when users aren’t active in the chat interface.
- The platform is designed to be tool-agnostic, targeting companies with mixed software stacks rather than those locked into a single vendor’s ecosystem.
- Enterprise security controls — including permission scopes and audit logs — are built in from day one.
- Availability is tied to ChatGPT Enterprise and Team plans, keeping this firmly in the B2B lane for now.
- The real adoption challenge won’t be technical — it’ll be getting teams to trust agents with consequential actions and building the governance structures to support that.
Frequently Asked Questions
What are workspace agents in ChatGPT?
Workspace agents are AI-powered automation agents built into ChatGPT and powered by OpenAI’s Codex model. They can perform complex, multi-step tasks across connected business tools, running in the cloud without requiring constant user input. Think of them as persistent AI workers that can be assigned ongoing workflows rather than one-off questions.
Who are workspace agents designed for?
They’re built primarily for teams and enterprises using ChatGPT Enterprise or ChatGPT Team plans. Businesses with repetitive cross-tool workflows — data processing, reporting, lead management, documentation — will find the most immediate value. Developers can also use them to build and deploy internal tooling with relatively low overhead.
How do workspace agents compare to Microsoft Copilot or Google Gemini agents?
Microsoft’s Copilot agents are deeply integrated with Microsoft 365 and work best within that stack. Google’s Gemini automation tools are similarly anchored to Google Workspace. OpenAI’s workspace agents are designed to be tool-agnostic, connecting to a broader range of platforms — which gives them an edge in mixed-stack environments but depends on the depth of available integrations.
Are workspace agents available now, and what do they cost?
OpenAI announced workspace agents on April 22, 2026, with availability tied to ChatGPT Enterprise and Team plans. Specific pricing for agent usage beyond the base plan subscription hasn’t been fully detailed, and enterprise customers should expect to negotiate based on usage volume. OpenAI has not announced a consumer-tier version at this time.
OpenAI is clearly betting that agentic AI — not just conversational AI — is where enterprise spending consolidates over the next two years. With Codex as the technical foundation and ChatGPT’s existing enterprise footprint as the distribution channel, workspace agents are a credible entry into that race. The real test comes when the first wave of large deployments hits the complexity ceiling and teams find out how gracefully these agents fail. That’s when we’ll know if this is a durable product or a well-packaged demo.
