Skip to content
MyAI Guide
Newsletter
Launchchatgpt

OpenAI launches Workspace Agents in ChatGPT, the successor to custom GPTs

By Harsh Desai

TL;DR

OpenAI launched Workspace Agents inside ChatGPT on April 23, 2026. Cloud-based agents that teams build once, run together across ChatGPT plus Slack, Gmail, and Salesforce, then improve over time. Research preview free for ChatGPT Business, Enterprise, Edu, and Teachers until May 6, then credit-based pricing. Framed as the evolution of custom GPTs.

What shipped

OpenAI rolled out Workspace Agents inside ChatGPT on April 23, 2026. This is OpenAI's most significant team-facing product launch of the quarter.

Workspace Agents let organizations build a custom agent once, share it across the company, and have it run autonomously across ChatGPT plus third-party tools like Slack, Gmail, and Salesforce.

How they work

Three behavioural commitments OpenAI called out in the launch blog:

  1. Gather context from the right systems. The agent pulls the information it needs from connected tools before acting.
  2. Follow team processes. Workflows are codified rather than improvised per-run.
  3. Ask for approval when needed. Human-in-the-loop gating for destructive or high-stakes actions.
  4. Keep work moving across tools. Cross-app orchestration, not just single-app automation.

Two examples in the launch materials:

  • •A product-feedback agent that searches the web for customer mentions and posts a weekly report to Slack.
  • •A sales agent that drafts follow-up emails in Gmail after a meeting.

Pricing

Free in research preview until May 6, 2026. After that, pricing shifts to a credit-based model. Available on ChatGPT Business, Enterprise, Edu, and Teachers plans. Individual ChatGPT Plus or Free users do not get access at launch.

What this replaces

OpenAI is framing Workspace Agents as the successor to custom GPTs (the chatbot builder launched in November 2023). The company's own statement: Workspace Agents are an "evolution" of GPTs, and GPTs stay available "while teams test workspace agents with their workflows." Soon, OpenAI says, converting a GPT into a Workspace Agent will be a one-click operation.

Technical substrate

Workspace Agents run on OpenAI's Codex platform. This is the same infrastructure OpenAI used for its coding agents and is presumably the reason the same agent framework can handle research-style tasks and write-action tasks on the same execution layer.

Competitive context

This ships into a crowded agent market:

  • •Anthropic Claude Cowork already offers cross-tool agent workflows with file-system access on the user's own machine.
  • •OpenClaw (formerly Clawdbot) went viral this year with its "AI that actually does things" framing; OpenClaw founder Peter Steinberger now works at OpenAI.
  • •Hermes Agent from Nous Research reaches users across 16 messaging platforms with a self-improving skill system.
  • •Google Deep Research Max on Gemini 3.1 Pro shipped last week.

Workspace Agents is OpenAI's positioning move toward teams and enterprises specifically, rather than competing directly on the personal-agent axis where Claude Cowork and OpenClaw have traction.

For developers

The backend uses AgentKit, which OpenAI introduced in October 2025. AgentKit includes Agent Builder, a Connector Registry, and ChatKit. Workspace Agents surfaces this SDK as a product surface inside ChatGPT rather than requiring developers to build on AgentKit directly.

Who this matters for

  • Vibe Builder: Not for you if you are on ChatGPT Plus or Free. Stay on Claude Cowork, Hermes Agent, or OpenClaw for personal workflows. Revisit if your team moves to ChatGPT Business.
  • Basic User: If your company uses ChatGPT Business or Enterprise, ask your admin to try Workspace Agents while it is free. Start with one repetitive workflow like weekly customer feedback reports.
  • Developer: Workspace Agents surfaces AgentKit as a product. If you were building against AgentKit directly, the shareable-team-agent abstraction is now done for you. Consider whether your custom AgentKit code is still the right layer.

What to watch next

This is the launch OpenAI had to make. Anthropic shipped Claude Cowork earlier this year, OpenClaw went viral in Q1, Hermes Agent now reaches 16 messaging platforms with a learning loop, and Google Deep Research Max landed last week. Every major player now has a team-level agent product. OpenAI was the conspicuous gap, and Workspace Agents closes it.

What makes this competitive rather than just catch-up is the enterprise plumbing. Shareable team agents with approval gates and Slack plus Salesforce plus Gmail connectors is a different product category from a personal agent that runs on your laptop. OpenAI is not trying to out-Cowork Anthropic on personal productivity. They are trying to become the default agent layer for teams that already use ChatGPT Business or Enterprise.

The free window until May 6 is the move that matters most for adoption. Two weeks of free enterprise access is enough for a lot of teams to integrate a Workspace Agent into one real workflow, get the dependency, and stay through the credit-based pricing flip. Classic enterprise playbook; well-timed.

For vibe builders reading this: Workspace Agents is not for individual Plus subscribers. You will keep using Claude Cowork, Hermes Agent, or OpenClaw on your own machine. But if you operate inside a ChatGPT Business or Enterprise workspace, this is the shortest path from "we have a repetitive workflow" to "we have an agent that does it" without spinning up a custom build.

The GPTs-are-evolving framing is the quieter signal. Custom GPTs were OpenAI's first serious product beyond chat, and they never found escape velocity. Workspace Agents is the second attempt at the same problem (make GenAI reusable and shareable) with two years of agent framework learning baked in. Whether this one sticks depends on whether the team-sharing model is actually how real organizations want to buy AI, which is genuinely an open question.

by Harsh Desai

Source:openai.com

About chatgpt

View the full chatgpt page →All chatgpt updates

More from chatgpt