OpenAI launched a brand new paradigm and product in the present day that’s prone to have large implications for enterprises looking for to undertake and management fleets of AI agent employees.
Referred to as “Workspace Brokers,” OpenAI’s new providing basically permits customers on its ChatGPT Enterprise ($20 per person per 30 days) and variably priced Enterprise, Edu and Academics subscription plans to design or choose from pre-existing agent templates that may tackle work duties throughout third-party apps and information sources together with Slack, Google Drive, Microsoft apps, Salesforce, Notion, Atlassian Rovo, and different common enterprise purposes.
Put merely: these brokers may be created and accessed from ChatGPT, however customers also can add them to third-party apps like Slack, talk with them throughout disparate channels, ask them to make use of data from the channel they’re in and different third-party instruments and apps, and the brokers will go off and do work like drafting emails to your complete crew, chosen members, or pull information and make shows.
Human customers can belief that the agent will handle all this complexity and full the duty as requested, even when the person who requested it leaves.
It is the top of “babysitting” brokers and the beginning of letting them go off and get shit carried out for what you are promoting — in accordance with your outlined enterprise processes and permissions, after all.
The product expertise seems centered on the Brokers tab within the ChatGPT sidebar, the place groups can uncover and handle shared brokers.
This features as a type of crew listing: a spot the place brokers constructed by coworkers may be reused throughout a workspace. The broader concept is that AI turns into much less of a person productiveness trick and extra of a shared organizational useful resource.
On this sense, OpenAI is concentrating on one in every of workplace work’s oldest ache factors: the handoff between individuals, techniques, and steps in a course of.
OpenAI says workspace brokers can be free for the subsequent two weeks, till Could 6, 2026, after which credit-based pricing will start. The corporate additionally says extra capabilities are on the way in which, together with new triggers to begin work mechanically, higher dashboards, extra methods for brokers to take motion throughout enterprise instruments, and assist for workspace brokers in its AI code technology app, Codex.
For extra data on the best way to get began constructing and utilizing them, OpenAI recommends heading over to its on-line academy web page on them right here and its assist desk documentation right here.
The Codex spine
Essentially the most important shift on this announcement is the transfer away from purely session-based interplay. Workspace brokers are powered by Codex — the cloud-based, partially open-source AI coding harness that OpenAI has been aggressively increasing in 2026 — which supplies them entry to a workspace for recordsdata, code, instruments, and reminiscence.
OpenAI says the brokers can do excess of reply a immediate. They will write or run code, use linked apps, bear in mind what they’ve realized, and proceed work throughout a number of steps.
That description traces up intently with the capabilities OpenAI shipped into Codex simply six days in the past, together with background laptop use, greater than 90 new plugins spanning instruments like Atlassian Rovo, CircleCI, GitLab, Microsoft Suite, Neon by Databricks, and Render, plus picture technology, persistent reminiscence, and the flexibility to schedule future work and get up by itself to proceed throughout days or perhaps weeks.
Workspace brokers inherit that plumbing. When one pulls a Friday metrics report, it’s successfully spinning up a Codex cloud session with the precise instruments hooked up, operating code to fetch and rework information, rendering charts, writing the narrative, and persisting what it realized for subsequent week.
When that very same agent is deployed to a Slack channel, it’s a Codex occasion listening for mentions and threading its work again in.
That is the technical determination enterprise consumers ought to deal with. Constructing an agent on a code-execution substrate fairly than a pure LLM-call-and-response loop is what offers workspace brokers the flexibility to do actual work — reworking a CSV, reconciling two techniques of file, producing a chart that’s really right — fairly than describing what the work would appear like.
Persistence and scheduling
In earlier AI assistant fashions, progress paused when the person stopped interacting. Workspace brokers change that by operating within the cloud and supporting long-running workflows. Groups also can set them to run on a schedule.
Which means a recurring reporting agent can pull information on a set cadence, generate charts and summaries, and share the outcomes with a crew with out anybody manually kicking off the method.
Right here at VentureBeat, we analyze story site visitors and person return price on a weekly foundation — precisely the type of recurring, multi-step, multi-source job that might theoretically be automated with a single workspace agent. Any enterprise with a weekly reporting rhythm pulling from dynamic information sources is prone to discover a use for these brokers.
Brokers additionally retain reminiscence throughout runs. OpenAI says they are often guided and corrected in dialog, in order that they enhance the extra a crew makes use of them.
Over time they begin to replicate how a crew really works — its processes, its requirements, its most popular methods of dealing with recurring jobs — which is a meaningfully completely different proposition from the static instruction-set GPTs that preceded them.
The built-in ecosystem
OpenAI’s declare is that brokers ought to collect data and take motion the place work already occurs, fairly than forcing groups right into a separate interface. That time turns into clearest within the Slack examples. OpenAI’s launch supplies present a product-feedback agent working inside a channel named #user-insights, answering a query about current mobile-app suggestions with a themed abstract pulled from a number of sources.
The corporate’s demo lineup walks by a pattern crew listing of brokers: Spark for lead qualification and follow-up, Slate for software-request assessment, Tally for metrics reporting, Scout for product suggestions routing, Trove for third-party vendor threat, and Angle for advertising and marketing and net content material.
OpenAI additionally shared extra practical examples its personal groups use internally — a Software program Reviewer that checks worker requests towards approved-tools coverage and recordsdata IT tickets; an accounting agent that prepares elements of month-end shut together with journal entries, balance-sheet reconciliations, and variance evaluation, with workpapers containing underlying inputs and management totals for assessment; and a Slack agent utilized by the product crew that solutions worker questions, hyperlinks related documentation, and recordsdata tickets when it surfaces a brand new challenge.
In a way, it’s a continuation of the philosophy OpenAI espoused for people with final week’s Codex desktop launch: the agent joins the workflow the place work is already taking place, attracts in context from the encompassing apps, takes motion the place permitted, and retains shifting.
From GPTs to a broader agent push
Workspace brokers should not a standalone launch. They sit inside a roughly 12-month arc by which OpenAI has been systematically rebuilding ChatGPT, the API, and the developer platform round brokers.
Workspace brokers are explicitly positioned by OpenAI as an evolution of its {custom} GPTs, launched in late 2023, which gave customers a strategy to create custom-made variations of ChatGPT for explicit roles and use instances.
Nevertheless, now OpenAI says it’s deprecating the {custom} GPT customary for organizations in a yet-to-be decided future date, and would require Enterprise, Enterprise, Edu and Academics customers to replace their GPTs to be new workspace brokers.
People who’ve made {custom} GPTs can proceed utilizing them for the foreseeable future, in accordance with our sources on the firm.
In October 2025, OpenAI launched AgentKit, a developer-focused suite that features Agent Builder, a Connector Registry, and ChatKit for constructing, deploying, and optimizing brokers.
In February 2026, it launched Frontier, an enterprise platform centered on serving to organizations handle AI coworkers with shared enterprise context, execution environments, analysis, and permissions.
Workspace brokers arrive because the no-code, in-product entry level that sits on high of that stack — even when OpenAI doesn’t explicitly describe the architectural relationship in its supplies.
The subtext throughout all three launches is identical: OpenAI has determined that the way forward for ChatGPT-for-work is fleets of permissioned brokers, not single chat home windows — and that GPTs, its first try at letting companies customise ChatGPT, weren’t sufficient.
Governance and enterprise safeguards
As a result of workspace brokers can act throughout enterprise techniques, OpenAI places heavy emphasis on governance. Admins can management who’s allowed to construct, run, and publish brokers, and which instruments, apps, and actions these brokers can attain.
The role-based controls are extra granular than those most custom-GPT rollouts ever had: admins can toggle, per function, whether or not members can browse and run brokers, whether or not they can construct them, whether or not they can publish to the workspace listing, and — individually — whether or not they can publish brokers that authenticate utilizing private credentials.
That final setting is the dangerous case, and OpenAI explicitly recommends protecting it narrowly scoped.
Authentication itself is available in two flavors, and the selection has actual penalties. In end-user account mode, every one that runs the agent authenticates with their very own credentials, so the agent solely ever sees what that particular person is allowed to see.
In agent-owned account mode, the agent makes use of a single shared connection so customers do not need to authenticate at run time. OpenAI’s documentation strongly recommends service accounts fairly than private accounts for the shared case, and flags the data-exfiltration threat of publishing an agent that authenticates as its creator.
Write actions — sending e mail, modifying a spreadsheet, posting a message, submitting a ticket — default to All the time ask, requiring human approval earlier than the agent executes.
Builders can chill out particular actions to “By no means ask” or configure a {custom} approval coverage, however the default posture is human-in-the-loop.
OpenAI additionally claims built-in safeguards towards prompt-injection assaults, the place malicious content material in a doc or net web page tries to hijack an agent. The declare is welcome however not but confirmed within the wild.
For organizations that need deeper visibility, OpenAI says its Compliance API surfaces each agent’s configuration, updates, and run historical past.
Admins can droop brokers on the fly, and OpenAI says an admin-console view of each agent constructed throughout the group, with utilization patterns and linked information sources, is coming quickly.
Two caveats value flagging for security-sensitive consumers: workspace brokers are off by default at launch for ChatGPT Enterprise workspaces pending admin enablement, and they aren’t out there in any respect to Enterprise clients utilizing Enterprise Key Administration (EKM).
Analytics and early buyer sign
OpenAI additionally ships an analytics dashboard aimed toward serving to groups perceive how their brokers are getting used. Screenshots within the launch supplies present measures like whole runs, distinctive customers, and an exercise feed of current runs, together with one by a person named Ethan Rowe finishing a run in a #b2b-sales channel.
The mockup element helps OpenAI’s broader level: the corporate needs organizations to measure not simply whether or not brokers exist, however whether or not they’re getting used.
The clearest early-adopter sign within the launch itself comes from Rippling. Ankur Bhatt, who leads AI Engineering on the HR platform, says workspace brokers shortened the normal growth cycle sufficient {that a} gross sales marketing consultant was capable of construct a gross sales agent with out an engineering crew. “It researches accounts, summarizes Gong calls, and posts deal briefs instantly into the crew’s Slack room,” Bhatt says. “What used to take reps 5–6 hours per week now runs mechanically within the background on each deal.”
OpenAI’s announcement names SoftBank Corp., Higher Mortgage, BBVA, and Hibob as further early testers.
The period of the digital coworker
Workspace brokers don’t land in a vacuum. They land in the course of a broader OpenAI push — by AgentKit, by Frontier, by the Codex overhaul — to make brokers extra persistent, extra linked, and extra helpful inside actual organizational workflows.
Additionally they land in a deeply crowded discipline: Microsoft Copilot Studio is wired into the Microsoft 365 base, Google is pushing Agentspace, Salesforce has rebuilt itself as agent infrastructure with Agentforce, and Anthropic just lately launched Claude Managed Brokers, all completely different flavors of comparable concepts — brokers that lower throughout your apps and instruments, take actions on schedules repeatedly as desired, and retain a point of reminiscence, context, and permissions and insurance policies.
However this launch issues as a result of it turns OpenAI’s technique into one thing concrete for the groups already paying for ChatGPT, and since it quietly retires the product these groups had been most just lately informed to standardize on.
If workspace brokers dwell as much as the pitch — shared, reusable, scheduled, permissioned coworkers that comply with accredited processes and maintain work shifting when their human is offline — it will mark a significant change in what office software program does. Much less passive software program ready for enter, extra energetic techniques serving to groups coordinate, execute, and transfer sooner collectively.
The period of the digital coworker has begun. And, on OpenAI’s plans a minimum of, the period of the {custom} GPT is ending.
