Quick Answer
Teams that repeat one workflow across shared tools now have a sensible reason to test ChatGPT workspace agents. OpenAI introduced them on 2026-04-22 in Introducing workspace agents in ChatGPT, and the current public docs point to a practical rollout path: build one narrow workflow, keep risky write actions on approval, preview it with realistic inputs, and only then add a schedule or Slack.
Regular ChatGPT chats still fit one-off drafting. Projects fit ongoing work with persistent context. Workspace agents fit work that is recurring, shared, tool-connected, and meant to keep running outside one conversation. If you need a refresher on the Project and Skill side of that split, start with our guide to ChatGPT Projects and Skills.
Before You Start
The launch post says workspace agents are in research preview for ChatGPT Business, Enterprise, Edu, and Teachers as of 2026-04-23. OpenAI's Help Center adds the operational caveats most tutorials skip: workspace agents are off by default at launch for Enterprise, they are not available for ChatGPT Enterprise customers with EKM, and Slack use has extra setup and authentication requirements. Those details come from OpenAI's launch post, Academy guide, and the Help Center article ChatGPT Workspace Agents for Enterprise and Business.
Check these requirements before you build anything:
| What to check | Why it matters |
|---|---|
| Eligible plan | Workspace agents are currently in research preview rather than broad general availability. |
| Workspace permissions | Enterprise admins can separately enable agent browsing, building, and publishing. |
| EKM status | Enterprise customers with EKM cannot use workspace agents right now. |
| Tool and app access | The builder only exposes tools and connectors enabled in your workspace, and some require authentication before the agent can run. |
| Slack readiness | If you want Slack, the ChatGPT Agents app must be added to the channel, Slack access may require admin approval, and every app connection used by that agent must use shared authentication. |
Most teams can reduce the preflight to four checks:
- Confirm your plan exposes workspace agents.
- Confirm your admins allow you to build or publish them.
- Confirm the workflow's apps are already connected.
- Decide whether the first version really needs Slack.
Unclear permissions or missing connectors derail more pilots than prompt quality does. Resolve those first.
Regular Chat vs Project vs Workspace Agent
OpenAI describes workspace agents as an evolution of GPTs that run in the cloud, can use tools, can be shared across an organization, and can work in ChatGPT or Slack. That makes them a different product shape from a regular chat or a Project.
| If the job looks like this | Best default |
|---|---|
| One-off drafting, summarizing, brainstorming | Regular ChatGPT chat |
| Ongoing work with files and instructions for one project | ChatGPT Project |
| Shared workflow with triggers, tools, approvals, and handoffs | ChatGPT workspace agent |
Regular chats handle one-off thinking. Projects hold ongoing context. Skills keep reusable instructions. Workspace agents run the repeated workflow across tools.
Step 1: Pick One Workflow That Is Actually Agent-Shaped
OpenAI's Academy guide says workspace agents work best when the job is repeatable, structured, time-based or event-driven, and tool-based. That filter matters more than the builder itself.
Start with jobs like these:
- every Friday, pull KPI inputs and draft a weekly report
- every weekday, review new product feedback and route it by theme
- when a lead form arrives, build an account brief and draft follow-up
- in a Slack channel, answer common internal questions and file a ticket when documentation is missing
Avoid these for version one:
- brand strategy
- open-ended research with unclear success criteria
- high-risk workflows without a clear approval point
- anything where the team cannot agree on what a good output looks like
For a first rollout, a weekly metrics reporter or feedback triage agent is usually safer than a broad cross-functional assistant. The scope is narrow, the review loop is fast, and the stop point is obvious.
Step 2: Create The Agent From Agents Or A Template
OpenAI's Help Center says you can create a workspace agent from Agents in the ChatGPT sidebar, either by selecting Create or by starting from Browse templates. The shortest reliable builder path is:
- Open Agents.
- Select Create.
- Enter a plain-language prompt describing the job, or start blank.
- Review the draft plan.
- Select Build this agent.
- Refine the workflow in the builder.
- Use Preview before you publish.
- Select Create when the run looks right.
Do not start by describing a role. Start by describing one bounded job.
A good builder prompt:
Build a workspace agent for our weekly growth report.
It should:
- run every Friday at 9 AM
- gather KPI inputs from approved tools
- draft a report with these sections: key changes, risks, wins, next actions
- post the draft to the team channel only after approval
If source data is missing, it should stop and tell me what is missing instead of guessing.
A weak builder prompt:
Help our team work better with AI.
The second prompt is too broad to turn into a stable workflow.
Step 3: Add Tools, Auth, And Skills Carefully
From the builder's Tools section, OpenAI says you can add apps, your own custom MCPs, and additional tools such as web search or image generation. The same Help Center article also says each app connection can use either end-user authentication or an agent-owned account.
Choose the auth model deliberately:
- use end-user auth when each runner should act as themselves
- use an agent-owned account when the workflow needs a shared connection
- use a service account when possible instead of a personal account
- grant the least privilege the workflow actually needs
If your workflow will run in Slack, this is not optional. OpenAI's Help Center says all app connections used by a Slack agent must use shared authentication. If you keep personal connections, the agent may work in ChatGPT but fail in Slack.
OpenAI's Academy guide also notes that workspace agents can use Skills inside their instructions. Use them for repeatable formatting or review logic, not for basic product setup.
Copyable planning template
Objective:
Create a weekly team report from approved source systems.
Trigger:
Every Friday at 9:00 AM.
Process:
1. Pull the latest KPI inputs.
2. Compare with the previous period.
3. Flag material changes.
4. Draft a short summary with 3 actions.
5. Prepare the draft for the team channel.
Tools:
Analytics source, shared document, Slack.
Guardrails:
- Do not edit dashboards.
- Do not message external recipients.
- Ask for approval before sending outside the team channel.
- If a source is missing or stale, stop and report the gap.
Stop there for version one. Extra exceptions usually make the first draft worse.
Step 4: Configure ChatGPT Access, Schedule, And Slack Only If Needed
Every workspace agent automatically has a ChatGPT entry point. In the Help Center setup flow, OpenAI says you can control whether the agent is private, shareable by link, or published to the organization directory from the ChatGPT channel page.
If the agent should run on a cadence, OpenAI says to add the schedule from the ChatGPT channel page:
- Select Add schedule.
- Choose the channel.
- Choose the schedule type and frequency.
- Add any extra instructions.
- Save the schedule.
Wait until the ChatGPT version is stable before adding Slack.
For Slack, OpenAI's docs say you need to:
- Add the ChatGPT Agents app to the Slack channel.
- Connect the Slack workspace to the agent.
- Create a Slack handle for the agent.
- Choose the Slack channel where it should operate.
- Decide whether it responds to every message or only when mentioned.
Two operational gotchas matter here:
- a Slack admin may need to approve access or related permissions before setup finishes
- if you change the Slack workspace later, OpenAI says the previous Slack channels are removed and the Slack connection has to be rebuilt
Validate the ChatGPT version first, then add Slack.
Step 5: Set Write Approvals And Stop Conditions Before Sharing
Once the workflow can touch real systems, approval design matters more than prompt polish.
OpenAI's launch post says teams decide what tools and data the agent can use, what actions it can take, and when it needs approval. The Help Center adds one important product detail: write actions are set to Always ask by default during an agent run, and some apps may also support Never ask or Custom approval.
For version one, keep the policy simple:
- reading approved internal sources: allowed
- drafting summaries and tickets: allowed
- editing records, sending messages, or posting outside the team channel: approval required
- missing or conflicting source data: stop and escalate
OpenAI uses similar examples for sensitive actions, including editing a spreadsheet, sending an email, or adding a calendar event.
Copyable approval block
Approvals:
- Ask before sending to any external recipient.
- Ask before posting outside the designated team channel.
- Ask before editing spreadsheets, calendars, CRM records, or tickets.
Stop conditions:
- Stop if more than one required source is missing.
- Stop if confidence is low and an owner cannot be inferred.
- Stop if the request would expose data outside the approved audience.
Step 6: Preview It With Clean Inputs And Messy Inputs
OpenAI's Academy guide says agent building should be iterative, and the Help Center says you should use Preview inside the builder before creating the agent. Preview is the cheapest place to catch a bad workflow.
Run two tests before you publish:
- Clean-input test: every source exists, formatting is normal, and the workflow should succeed.
- Messy-input test: one source is missing, one field is ambiguous, or one connector returns incomplete data.
Use prompts like these:
Run the workflow using last week's normal input shape and show me the final draft plus any approvals you would request.
Now simulate the same workflow with one missing source and explain whether you should stop, continue with a partial draft, or ask for clarification.
When the agent guesses instead of pausing, tighten the instructions before you share it.
Copyable Workflow 1: Weekly Metrics Reporter
Weekly metrics reporting is a good first deployment for most ops, growth, and marketing teams.
Builder prompt
Build a workspace agent for our weekly growth report.
It should:
- run every Friday at 9 AM
- gather KPI inputs from approved tools
- compare the latest period with the previous one
- draft a one-page report with these sections: key changes, risks, wins, next actions
- prepare the result for the team channel
- require approval before any broader distribution
If source data is missing, stale, or contradictory, it should stop and list the gap instead of guessing.
Output format
Weekly Growth Report
- Top 3 changes
- Metrics that moved materially
- Risks or data gaps
- Recommended next actions
Why it works as a first rollout
- the trigger is obvious
- the output format is easy to review
- the workflow touches real systems
- the approval boundary is clear
Copyable Workflow 2: Feedback Triage Agent
Feedback triage is a strong second pattern for product, support, and community teams.
Builder prompt
Build a workspace agent that reviews new product feedback from our approved Slack and support inputs.
It should:
- run every weekday at 8 AM
- group related feedback into themes
- mark urgency when multiple users report the same blocker
- draft a short triage summary
- suggest the right owner or team
- create only draft tickets unless a human approves submission
It must not invent missing details. If a report is too vague to route, label it Needs Clarification.
Output format
Daily Feedback Triage
Theme:
Evidence:
Suggested owner:
Suggested action:
Confidence:
Skip this pattern when
- when feedback sources are mostly private messages your team cannot legally centralize
- when there is no routing taxonomy yet
- when the real problem is missing product ownership rather than missing triage labor
Common Mistakes
1. Starting with a role instead of a workflow
"Build us an AI operations manager" is not a workflow. "Every weekday, summarize support escalations and draft owner assignments" is.
2. Treating Slack as the default first surface
Slack is useful, but it adds more configuration, authentication, and channel-governance work. Validate the agent in ChatGPT first.
3. Skipping stop conditions
An agent without clear stop conditions can look convincing in a clean demo and fall apart on the first messy run.
4. Making version one too broad
One trigger, one output shape, one approval policy. That is enough for the first release.
Availability And Pricing
As of 2026-04-23, OpenAI says workspace agents are available in research preview for ChatGPT Business, Enterprise, Edu, and Teachers. The current Help Center also says the feature is off by default at launch for Enterprise and not available for ChatGPT Enterprise customers with EKM.
Three rollout details matter:
- GPTs remain available while teams test workspace agents
- workspace agents are free until 2026-05-06
- OpenAI says credit-based pricing starts after that date
The free window makes it easier to validate one narrow workflow before you commit to broader rollout.
When To Start
Start now if your team already has a repetitive workflow with a clear owner, a clear trigger, and a clear approval boundary.
Wait if your source systems are still messy, if admins have not opened the required access, or if nobody has agreed on what the output should look like. Workspace agents are a good fit for operational repetition, not for unresolved process design.
First Release Checklist
Before you publish version one, confirm these five items:
- one workflow, one trigger, and one output format are defined
- the required apps are connected with the right auth model
- risky write actions still require approval
- Preview succeeds once with clean input and once with messy input
- the team knows where the agent runs first: ChatGPT only, or ChatGPT plus Slack
FAQ
Who should use ChatGPT workspace agents instead of a normal ChatGPT chat?
Use a workspace agent when the work is shared, repeatable, tool-based, and needs a stable output format. Use a regular chat for one-off drafting or brainstorming.
What do I need before building a workspace agent?
You need an eligible ChatGPT workspace plan, admin permission to browse or build agents if your workspace restricts them, and the apps or connectors the workflow depends on. If you want to run the agent in Slack, you also need Slack approval and shared app authentication.
Can ChatGPT workspace agents run on a schedule and in Slack?
Yes. As of 2026-04-23, OpenAI says workspace agents can run on schedules and can be used in ChatGPT and Slack. For Slack, the ChatGPT Agents app must be added to the channel, and the agent's app connections must use shared authentication.
Do workspace agents replace GPTs and Skills?
Not yet. OpenAI describes workspace agents as an evolution of GPTs, says GPTs remain available while teams test them, and notes that workspace agents can use Skills in their instructions.
Verification Note
Reviewed on 2026-04-24.
Reconfirmed on 2026-04-24 against OpenAI's launch post and Academy guide: launch date, current availability, research preview status, Skills support, repeatable-workflow guidance, and the free-until-2026-05-06 pricing window. Help Center operational details in this article, including Enterprise default-off, EKM restrictions, builder flow, scheduling flow, Slack setup, shared-auth requirements, and write-action approval defaults, were last checked on 2026-04-23.
Official sources: Introducing workspace agents in ChatGPT, Workspace agents, and ChatGPT Workspace Agents for Enterprise and Business.