ChatGPT Workspace Agents: Codex Team Automation Guide (2026)

On April 22, 2026, OpenAI launched workspace agents in ChatGPT — and the difference between this and previous GPTs is not cosmetic. Where GPTs were essentially saved system prompts with a handful of connected tools, workspace agents are cloud-running processes that can be scheduled, deployed into Slack, write and execute code, remember context across sessions, and operate on your team’s data without you being present. If you have been waiting for ChatGPT to feel less like a chatbot and more like a capable teammate, this is the release that changes that calculus.

This guide covers everything your team needs to know: what workspace agents actually are, how they differ from GPTs, how to set one up in under ten minutes, which integrations matter most, and what the pricing looks like after the free preview ends on May 6.

Try it yourself: Free Coding Assistant ROI Calculator — free, no signup, runs in your browser.

What Are Workspace Agents? (And Why They Are Not Just Better GPTs)

Workspace agents are OpenAI’s evolution of GPTs — but calling them “better GPTs” undersells the architectural shift. GPTs were conversational templates: you gave ChatGPT a system prompt, maybe connected a few tools like web browsing or DALL·E, and saved it for reuse. The interaction model was synchronous — you prompt, it responds, done.

Workspace agents are built for asynchronous, multi-step work. They run in the cloud so they keep working after you close the browser. They can be triggered on a schedule. They can receive requests from Slack and respond autonomously. They retain memory across sessions — what they learned from last week’s run informs this week’s. And critically, they are shared at the workspace level, meaning a team can build one agent and every member benefits from it.

The underlying engine is Codex — the same model powering OpenAI’s developer coding assistant — which means workspace agents can not only plan and reason about multi-step tasks but actively write and run code as part of completing them. An agent tasked with preparing a weekly analytics summary can query your connected data sources, write a Python script to process the output, and deliver a formatted report, all without your intervention.

How Workspace Agents Work: The Architecture

Understanding the architecture helps you build better agents. Each workspace agent consists of three layers working together.

Instructions and Persistent Memory

The agent’s core identity: what it does, what tone it takes, what it should and should not attempt. This is similar to a GPT system prompt but informed by a persistent memory layer — the agent updates its own understanding of your workflows over time. If you correct its report format in week two, it remembers that adjustment in week three without you re-prompting.

Connected Tools and Apps

Workspace agents connect to over 60 apps directly from ChatGPT Business workspaces, including Slack, Google Drive, SharePoint, GitHub, Jira, Atlassian, Notion, Microsoft 365, and Salesforce. Each integration grants the agent specific permissions — read-only versus read-write, specific channels or folders — that workspace admins configure through the ChatGPT admin console.

Beyond pre-built integrations, agents can write and execute code (primarily Python), browse the web, and call external APIs where webhooks are configured. This means an agent can pull data from a proprietary internal system through an API call, process it with a Python script, and push the result to a Slack channel — in a single autonomous run.

Triggers

Three trigger modes are available at launch:

• Manual: A team member opens the agent in ChatGPT or Slack and asks it to do something.
• Scheduled: The agent runs automatically at configured intervals — daily, weekly, on the first Monday of the month. Ideal for recurring reports, status digests, and periodic data pulls.
• Slack-triggered: The agent is deployed into a Slack channel or DM and picks up requests as they arrive, responding without requiring a browser session from any team member.