ChatGPT Workspace Agents just unlocked something I've been waiting 18 months for.
Multi-agent collaboration with a simple @-mention.
One agent hands off work to another agent.
That agent hands off to a third.
Your entire workflow runs in a chain of specialised AI workers — while you sleep.
This is the feature nobody on Twitter is talking about.
And it's going to change how teams build AI systems in 2026.
Here's exactly how it works.
What Are ChatGPT Workspace Agents?
Workspace agents are codex-powered AI agents that run 24/7 in the cloud.
They're built by describing the job in plain English.
No code.
No scripts.
They live in a new AI agents section in the ChatGPT sidebar.
Available on business, enterprise, edu or teacher plans.
Free until May 6th.
The killer feature for teams though isn't the solo agent.
It's the fact that agents can talk to each other.
The @-Mention That Changes Everything
Inside any workspace agent chat, you can type @ and pick another agent.
That agent now has context on what the current agent is working on.
It picks up the task.
It runs it.
It can @-mention another agent to continue the chain.
This is how human teams work.
Researcher passes brief to writer.
Writer passes draft to editor.
Editor passes to publisher.
Workspace agents just made that AI-native.
See how I'm wiring multi-agent chains inside AI Profit Boardroom →
A Real Multi-Agent Chain I Built
Let me walk you through one of my chains.
Agent 1: Research Agent
Job: Pull the top 5 AI news stories every morning at 7am.
Tools: Web search, Google Drive.
Output: A brief document saved to Drive.
Handoff: @-mentions the writer agent with the brief link.
Agent 2: Writer Agent
Job: Turn the brief into a 1500-word SEO-optimised article.
Tools: Google Drive, memory of my writing style.
Output: A Google Doc draft.
Handoff: @-mentions the editor agent with the draft.
Agent 3: Editor Agent
Job: Tighten the copy, check UK grammar, add internal links.
Tools: Google Drive, Notion (for my link index).
Output: A publish-ready doc.
Handoff: @-mentions the distribution agent.
Agent 4: Distribution Agent
Job: Post to my CMS, schedule social posts, email my list.
Tools: Gmail, Wix, HubSpot.
Output: Published article, scheduled social, email in queue.
Handoff: Notifies me in Slack when done.
Every morning I wake up to a published article.
I write zero lines of code to orchestrate this.
Every handoff is an @-mention.
Why @-Mention Handoff Beats Traditional Workflows
Traditional automation tools like Zapier or n8n make you wire every connection by hand.
One node fails, the whole flow dies.
Workspace agents with @-mention handoff are resilient.
Each agent has its own memory.
Each agent has its own instructions.
Each agent has its own approval gates.
When one agent hits a wall, it flags the issue instead of silently failing.
And because the handoff is conversational, you can jump in mid-chain and redirect.
"Actually, use the competitor angle instead."
The chain adapts.
How to Set Up Multi-Agent Handoff
Here's the exact process.
Step 1: Build your first agent. Describe the job in plain English. ChatGPT generates the plan with name, channels, capabilities, skills and memory.
Step 2: Build your second agent. Same process. Give it a clear single-purpose job.
Step 3: In agent 1's instructions, tell it explicitly: "When the brief is ready, @-mention the writer agent with the brief link."
Step 4: In agent 2's instructions, tell it how to respond when @-mentioned. "When @-mentioned with a brief, draft a 1500-word article in the style of my saved examples."
Step 5: Test the handoff manually once. Trigger agent 1. Watch it @-mention agent 2. Confirm the handoff.
Step 6: Schedule agent 1 to run daily.
Done.
Your chain is now running 24/7.
The Apps Workspace Agents Can Use
Each agent in your chain can connect to these apps.
- Google Drive
- HubSpot
- Notion
- GitHub
- Gmail
- Google Sheets
- Wix
- Slack (the only non-ChatGPT channel currently supported)
WordPress isn't native yet.
You can build a custom connection if you're technical.
Or wait — OpenAI has said more integrations are in the pipeline.
Admin Controls for Multi-Agent Systems
The reason this is business-only becomes obvious once you build a chain.
Admins decide who can build agents.
Admins decide who can use which agents.
Admins decide which tools each agent can touch.
Approval gates can be placed at any handoff.
So your distribution agent can't publish until a human approves the final doc.
Audit logs track every @-mention and every tool call.
This is what makes multi-agent systems safe for teams of 50 or 500.
Common Mistakes I See
Mistake 1: Too many agents. Start with two. Get the handoff working. Add a third only when the first chain is solid.
Mistake 2: Overlapping jobs. Each agent needs one clear role. "Researcher + writer" confuses the handoff. Split them.
Mistake 3: No memory. Upload files to each agent so it knows your style, your brand, your examples. An agent without memory is a generic tool.
Mistake 4: No approval gate. Put a human in the loop for anything that touches a customer. Audit logs don't prevent bad emails going out — approval gates do.
Mistake 5: Skipping the preview run. Always run a preview before going live on a schedule. You'll catch 80% of bugs.
The Templates That Already Use Handoff
OpenAI ships 5 templates and three of them are designed for multi-agent patterns.
Product feedback routing agent — captures Slack feedback and hands off to the right product owner.
Lead outreach agent — qualifies leads and hands off to a CRM-updating agent.
Software review agent — triages requests, routes approvals, opens tickets in a handoff chain.
Start with a template, tear it apart, rebuild it your way.
What's Coming Next
OpenAI has confirmed the roadmap.
- New triggers — agents will start work automatically based on events, not just schedules.
- Better dashboards — see agent activity at a glance.
- More tool integrations — the current list will grow fast.
- Codex app support — tighter connection between workspace agents and the Codex developer tools.
The teams that build multi-agent chains now are going to have a compounding advantage by Q3.
Related Reading
FAQ
Q: How does @-mention handoff work in ChatGPT workspace agents?
Inside any agent chat you type @ and pick another agent. That agent gets context on the current task and picks it up.
Q: Can a workspace agent @-mention more than one agent in sequence?
Yes. Each agent in the chain can @-mention the next one. You can build chains of 3-5 agents easily.
Q: Do multi-agent chains need coding?
No. You describe each agent's job in plain English. The @-mention is just typing @ in the instructions.
Q: What happens if one agent in the chain fails?
The agent flags the issue instead of silently failing. Audit logs show you exactly where the chain broke.
Q: Can I put an approval gate between agents?
Yes. You can require human approval at any handoff. This is especially useful before anything customer-facing.
Q: Is multi-agent handoff available on the ChatGPT personal plan?
No. Workspace agents require a business, enterprise, edu or teacher plan. Free until May 6th.