How to Use ChatGPT Workspace Agents (Step‑by‑Step Guide)

ChatGPT Workspace Agents are finally here—and they’re much more than just smarter chats. Agents can follow detailed instructions, use skills, connect to your apps, and actually complete work for you in the background.

In this guide, you’ll learn how to access Workspace Agents, build your first one, test it properly, and see real examples for sales, reporting, product feedback, and vendor risk management.

What ChatGPT Workspace Agents Actually Do

A ChatGPT Workspace Agent is a system that:

• Follows your instructions
• Uses skills (reusable sets of instructions)
• Works across tools and apps you connect
• Performs actions and produces real outputs

For example, you can ask an agent to handle email follow-ups from your CRM. The agent can:

• Read a new message or support ticket
• Summarize the request
• Draft a follow-up email
• Prepare a CRM note
• Wait for your review before sending anything

Behind the scenes, it might load skills like “follow-up writer” and “customer tone,” then use tools like CRM, email, and docs to do the work. The result: ready-to-review drafts and notes directly in ChatGPT.

How to Access the Agents Workspace

Workspace Agents live in the business side of ChatGPT, not the standard personal interface. To access them:

• Upgrade your ChatGPT plan to a business/workspace plan.
• Switch from your personal area to the business workspace.
• Open the sidebar and scroll until you see the Agents section.
• Click Browse agents to see agents you can use or adapt.

Once you’re in the workspace, you’ll see an Agents tab where you can create, edit, and monitor your agents.

Building Your First ChatGPT Agent

You don’t need to be technical to build an agent. The builder is conversational and does most of the heavy lifting for you.

1. Start a New Agent

• Click Agents in your sidebar.
• Click Create agent.
• Either pick a template or describe in plain language what you want the agent to do.

You can even use the dictate button to speak your instructions instead of typing.

2. Let the Agent Builder Draft the Plan

The agent builder will:

• Think through the best way to structure your agent
• Draft a plan and initial instructions
• Suggest tools, skills, and schedules if needed

The more context you provide—what data it should use, what outputs you expect, how often it should run—the better the first draft will be.

You’ll see a preview of the plan. From there you can:

• Ask for edits in natural language
• Accept the draft and click Start building

3. Configure Instructions, Tools, and Schedule

Once you start building, the screen splits:

• Left side: your chat with the agent builder (where you give feedback and ask for changes)
• Right side: the agent’s instructions, tools, skills, and configuration being built in real time

Here you can:

• Refine the agent’s core instructions (what it should and shouldn’t do)
• Connect apps (email, CRM, Slack, Google Drive, Linear, etc.)
• Set a schedule (e.g., daily, weekly, or on-demand)

If the agent needs you to authenticate into an app, you’ll be prompted to do that directly in the chat.

4. Save and Preview the Agent

When you’re happy with the setup, you can:

• Save the agent
• Use Preview to run a test workflow

Preview mode lets you watch the agent step through its instructions in real time and see the final outputs before you share it with your team.

Using and Managing Shared Agents

You can also use agents that others in your workspace have created.

To explore shared agents:

• Go to Browse agents in the sidebar.
• Click an agent to see its details.

On the details page, you can:

• Check which apps it needs to run (for example, Gmail, Slack, CRM)
• Connect your own accounts to those apps if required
• Click View agent to read its instructions
• Duplicate the agent if you want a version you can customize

Once you understand how an agent works, you can trigger a workflow using one of its sample prompts or your own request and watch it execute step by step.

How to Properly Test and Improve Your Agent

Before you rely on an agent or share it widely, you should test it with evals (evaluation tests). This helps you catch issues early and improve reliability.

1. Set Up Evals (Test Scenarios)

Create a set of test inputs and ask yourself three questions:

• Does it follow instructions?
• Does it produce a useful output?
• Does it respect guardrails? (e.g., not sending emails without approval)

Use three types of test inputs:

• Realistic inputs: Typical requests that reflect real work.
• Messy inputs: Incomplete, unstructured, or conflicting requests.
• Edge cases: Unusual but possible scenarios.

If quality drops on messy or edge cases, you’ve found a weak spot to fix.

2. Patch, Retest, and Iterate

When something goes wrong, describe the failure clearly in chat. For example:

“When I paste rough notes, you miss key details and ignore the format I asked for.”

Let the agent builder update the instructions based on your feedback, then run the same eval again. Reusing the same eval set lets you compare changes instead of guessing whether it’s actually better.

3. Use Preview for On-Demand Testing

From the agent builder, click Preview and add a prompt to manually trigger the agent. You’ll be able to:

• Watch the agent’s logic and tool calls in real time
• See where it might be overcomplicating or misunderstanding steps
• Leave feedback in the left sidebar so the builder can refine instructions

Running multiple scenarios before sharing the agent helps you avoid surprises when others start using it.

Real-World Examples of ChatGPT Workspace Agents

Here are some concrete examples of what Workspace Agents can do across different teams.

1. Product Feedback Analysis Agent

This agent reads product feedback, finds recurring issues, and routes work to the right team.

Workflow:

• Connects to web forums and other feedback sources using web search
• Connects to Slack to post summaries
• Groups feedback into themes and pain points
• Posts a daily summary to a product leadership Slack channel
• Creates or updates issues in Linear (or another ticket system)

When it runs, the agent:

• Reads new feedback from all connected sources
• Clusters similar problems
• Generates a clear summary for leadership
• Enriches existing tickets with new data or creates new tickets with full context

2. Weekly Reporting Agent

This agent automates a recurring weekly metrics report for your team.

Setup:

• Connect Google Drive so the agent can access spreadsheets and files
• Set the connection as agent-owned so it can run on a schedule without relying on one person’s login
• Ask ChatGPT to improve the workflow and make it more reliable

The builder might suggest creating a metrics calculation skill that encodes:

• Which metrics matter
• How to calculate and interpret them
• How the weekly readout should be structured

Then you can:

• Schedule the agent to run every Friday
• Give it a simple starting message like “Run analysis”

Each week, the agent will:

• Read the latest data from your spreadsheet
• Run code to calculate metrics and generate charts
• Compile everything into a shareable readout
• Log its activity so you can review every step it took

3. Sales Lead Handling Agent (Spark)

This example agent, called Spark, helps an SMB sales team work new inbound leads faster with personalized messaging.

You can ask it in natural language to:

• Research each new lead
• Score or grade the lead based on your qualification criteria
• Send the first outreach email
• Draft a follow-up email
• Schedule reminders so no touchpoint is missed

Behind the scenes, the builder:

• Attaches tools like Gmail and web search
• Builds detailed instructions for the workflow
• Asks you for any missing configuration

Once created and scheduled, Spark can:

• Pick up new contact form submissions
• Pull lead details and research the company
• Send a tailored first email
• Stage a follow-up in Gmail
• Set reminders for future outreach

This kind of sales agent pairs well with broader automation setups. If you’re interested in going deeper on multi-agent workflows, check out this guide on how to set up your first AI agent step by step.

4. Vendor Risk Assessment Agent (Trove)

Trove is a third-party risk manager similar to what a finance team might use for vendor due diligence.

Build process:

• Start with a prompt describing the vendor review workflow
• Specify the tools and systems it should use (e.g., internal docs, risk systems)
• Attach an existing risk assessment skill that encodes the team’s best practices

The builder turns this into a structured plan and configures tools, skills, and apps. No engineering resources are required.

When you run Trove in preview, you can:

• Watch all run traces and tool calls
• See how it gathers evidence on a vendor
• Observe how it applies the risk assessment skill
• Review the final structured report it generates for a human analyst

This cuts down the manual, time-consuming parts of vendor reviews while keeping human oversight for final decisions.

Next Steps

ChatGPT Workspace Agents turn plain-language instructions into end-to-end workflows that run across your tools and data. Start small with a single process—like weekly reporting or email follow-ups—then iterate with evals until you trust the results.

As you get comfortable, you can consolidate even more of your AI workflows into one place. If you’re exploring broader workspaces that unify multiple models and tools, you may also like this deep dive on using different AI models in a single workspace.