OpenAI turns GPTs into Codex-backed workspace agents you describe in plain language

The build surface is a description box, not a config screen

OpenAI is positioning workspace agents as an evolution of GPTs, but the entry point has changed. You click Agents in the ChatGPT sidebar, describe a workflow your team does often, and ChatGPT walks through defining the steps, connecting tools, adding skills, and testing until it behaves as expected.

That framing — describe the job or drop in a file — moves the construction of an agent out of a settings panel and into conversation. The announcement pairs this with templates for finance, sales, and marketing that ship with built-in skills and suggested tools, so the starting point is a working draft rather than a blank form.

The distinction from a GPT is concrete: these agents run in the cloud on Codex with access to files, code, tools, and memory, and they keep working when the creator is away, either on a schedule or by picking up requests in Slack.

Two surfaces where the work already happens

OpenAI is explicit that agents live in ChatGPT and Slack today, with more surfaces coming. The Slack deployment is the more telling one: OpenAI's own product team built an agent that answers employee questions in channels, links documentation, and files a ticket when it finds a new issue.

That is a deliberate choice to embed the agent in an existing conversation thread rather than a separate app. The accounting month-end close agent follows the same pattern — usable in ChatGPT for individuals, or added to Slack channels so a team can ask it questions and discuss its output in place.

For frontend teams, the interface is the coordination point: agents join the tools and conversations where work happens instead of asking people to switch contexts to a dedicated dashboard.

OpenAI's own claim: the model was never the hard part

The customer quote in the announcement is unusually direct about what this product is actually solving, and it is not model quality.

The hard part of building an agent is not the model. It's the integrations, memory, the user experience. Workspace agents collapsed that work, so one of our Sales Consultants built, evaluated, and iterated a Sales Opportunity agent end to end without an engineering team.Montana Labs

Ankur Bhatt of Rippling frames the win as a non-engineer shipping and iterating a working agent — one that researches accounts, summarizes Gong calls, and posts deal briefs to Slack. The stated payoff is reclaiming 5–6 hours a week per rep. Whether that generalizes past a research preview is the open question, but the intended buyer is clear: the person who has the workflow knowledge but not the integration work.

Control lives in the same interface as the build

The governance story is knit into the same surfaces. Creators decide which tools and data an agent can touch and can require approval before sensitive steps like editing a spreadsheet, sending an email, or adding a calendar event. After sharing, analytics show run counts and how many people use each agent.

For Enterprise and Edu, admins get role-based controls over who can use, build, and share agents, and control over which connected tools user groups can access. The Compliance API exposes every agent's configuration, updates, and runs, and admins can suspend an agent. OpenAI also cites built-in safeguards against prompt injection when agents hit misleading external content.

The implication: adoption now turns on the describe-to-agent flow, not Codex

Workspace agents are free until May 6, 2026, then move to credit-based pricing — so the current window is about getting teams to build. The pitch treats scattered team knowledge as the asset and the guided build flow as the mechanism to make it reusable: build once, correct it in conversation, then share or duplicate.

The bet worth watching is whether a described workflow reliably becomes a dependable agent for non-engineers, or whether the integration and permission work simply moves from code into a longer conversation with the builder. OpenAI's roadmap — triggers, better performance dashboards, and Codex-app support — suggests they expect the build-and-tune loop, not the model, to be where teams spend their effort.