OpenAI Workspace Agent: Build a Cloud-Based Digital Coworker Using Plain English

Source: OpenAI | Published: 2026-04-29T20:59:11Z

OpenAI launches Workspace Agent, letting business users build AI workers that execute cross-system tasks and run continuously in the cloud—no coding required, just natural language descriptions of their workflows.

Not a Chat Assistant — a Digital Colleague Running in the Cloud

Workspace Agents are fundamentally different from regular ChatGPT conversations. Powered by Codex, they can access files, code, and various tools to handle complex tasks spanning multiple systems — and they keep running in the cloud after you close your laptop. They can run on a schedule or live in a Slack channel to respond to team requests.

OpenAI drew clear lines between its agent products: Workspace Agents are for team collaboration, with shared access and cross-system orchestration; Codex is a personal agent for individual developers; and the Agents SDK is for dev teams that need to embed agents into their own products.

Building an Agent from Scratch with Natural Language

In the demo, solutions engineer Hojun built a sales meeting-prep agent starting from a blank page. No code whatsoever — type a prompt describing the agent's purpose, specify the tools it needs (Google Calendar, Google Drive, Gmail), and ChatGPT generates a plan, then auto-configures everything: connecting apps, setting permissions, and writing the instruction set.

Builders can give ChatGPT natural-language feedback at any point to adjust the agent's behavior. Hojun emphasized that this lets business experts — not IT or engineering teams — build and maintain these workflows themselves.

Meeting Prep: From Hours Every Evening to One Email Each Morning

Hojun juggles a packed calendar of client meetings. His agent automatically checks his calendar each morning, pulls client materials from Google Drive, supplements them with web searches, then generates a formatted briefing doc for each meeting — complete with an executive summary, client snapshot, meeting objectives, and more — and emails it to him.

"I used to spend hours every night putting together meeting prep docs… now the agent handles all of it in the background."

The briefing is mobile-friendly for a quick scan during the commute. Team members can also view the same briefing via a shared link, so everyone walks into the meeting aligned on the agenda and objectives.

Permission Granularity: Read Yes, Write No

When configuring the agent's tool permissions, Hojun walked through a key detail: connecting Google Calendar with only read access, all write operations turned off. The agent can view the schedule but can't modify or create calendar events. Gmail permissions were similarly scoped — only the ability to send emails, with unnecessary operations disabled.

This granular control is a core design principle of Workspace Agents: the builder always controls exactly what the agent can and cannot do.

Skills: Turning Team Best Practices into Agent Capabilities

Beyond connecting external apps, agents support "Skills." A skill is essentially a codified workflow or best practice — you can import existing processes from other platforms or have ChatGPT generate new ones on the spot.

For the meeting-prep agent, Hojun loaded a document formatting skill so briefings would come out with consistent tables, headings, and bullet-point structure. In the second demo — a software approval agent — he imported an evaluation workflow skill written by the procurement team, letting the agent apply the company's existing approval criteria.

Think of it this way: skills are the script, the agent is the actor. A single agent can use multiple skills, invoking different workflows depending on the task at hand.

Software Approval Agent: An IT Colleague Living in a Slack Channel

The second demo tackled a scenario familiar to nearly every company. There's always a Slack channel where employees request software tools and IT has to process each one manually. Hojun handed this workflow to an agent.

When an employee posts "I need Screen Studio" in Slack, the agent kicks off automatically: it queries the company's approved software list and usage data, researches Screen Studio's capabilities, then compares it against existing tools. In the demo, the agent found that the company already had a similar tool called Bloom, but all its licenses were fully allocated — so it escalated the request to IT and automatically created a Jira ticket with its full reasoning and recommendation attached.

Employees no longer need to research tools, write justifications, or file tickets themselves — the agent handles the entire chain end to end.

Hojun mentioned this agent is already running in production inside OpenAI, saving IT and procurement teams significant time while reducing redundant software spend.

Memory System: Agents Get Smarter Over Time

Every Workspace Agent can have memory enabled — a persistent file system where the agent saves notes, context, and outputs after each run, then draws on them in subsequent runs.

A notable design choice: memory is isolated by channel. In ChatGPT, each user gets an independent memory space; in Slack, each channel has its own memory shared across all messages in that channel. This means the same agent can accumulate different context in different settings.

The Next Chapter for GPTs

For teams already using custom GPTs, Workspace Agents are essentially GPTs evolved. OpenAI engineer Christina acknowledged that when GPTs first launched, model capabilities and platform infrastructure weren't ready to support truly powerful automation. Workspace Agents fill those gaps with multi-step workflows, cross-tool orchestration, scheduled triggers, and approval flows.

OpenAI will soon ship a migration tool to automatically convert GPTs into Workspace Agents, making it easy for teams to carry over existing workflows.

Pricing: Free Until May 6, Credit-Based After

Workspace Agents are in research preview and completely free until May 6. After that, they'll move to credit-based billing, with costs based on agent complexity and task volume — similar to how Codex billing works. Specific pricing details will be announced closer to the date.

Enterprise Controls: Who Builds, Who Uses, Who Sees

For Enterprise and Edu plans, admins can use role-based access controls to determine who can create agents, who can use them, and which apps agents are allowed to access. A compliance API provides monitoring and management capabilities for agent usage.

Every agent run generates a complete execution trace. Admins can review historical runs on the agent page, seeing exactly what each step did and what data it accessed. These records are also exportable via API. Currently only the agent's creator can edit it, but collaborative editing is on the way.