Build Hour: Workspace agents in ChatGPT

Key Concepts

• Workspace Agents: AI agents powered by OpenAI’s technology, designed for teams to handle complex, long-running, multi-system workflows.
• Codex: The underlying engine enabling agents to execute tasks in the cloud, even when the user's computer is offline.
• Skills: Codified best practices, processes, or scripts that an agent can utilize to perform specific tasks consistently.
• Memory: A persistent file system allowing agents to store notes and context, enabling them to improve and adapt over time.
• Agent Trace: A detailed audit log showing the "chain of thought" and specific actions taken by an agent during a run.
• No-Code/Natural Language Building: The ability to configure and iterate on agents using conversational prompts rather than traditional programming.

---

1. Overview of Workspace Agents

Workspace agents are designed for collaborative team environments. Unlike personal agents, they are built for shared work, can operate autonomously in the cloud, and integrate with enterprise tools. They are currently available in research preview for ChatGPT Business, Enterprise, and EDU plans.

• Key Differentiators:
  • Team-Oriented: Can be shared across departments.
  • Cloud-Native: Tasks continue to run even when the user is offline.
  • Extensible: Can access files, code, and external tools (e.g., Google Drive, Slack, Jira).

2. Practical Use Cases

The session highlighted two primary applications:

• Meeting Prep Agent: Automatically scans calendars, researches customer data in Google Drive/Web, and generates a formatted meeting brief (including executive readouts and objectives) before the workday begins.
• Software Review Agent: Operates within Slack to handle employee software requests. It researches vendor capabilities, checks against approved software stacks, evaluates license utilization, and can escalate requests to Jira if human intervention is required.

3. Building and Deployment Methodology

The process for building an agent is designed to be accessible to subject matter experts without requiring deep engineering support:

1. Initialization: Start from a blank slate or a template using natural language prompts.
2. Configuration: Define the agent’s scope by selecting specific apps and tools. Users can toggle "read" vs. "write" permissions to ensure security.
3. Skill Integration: Import existing organizational best practices or have ChatGPT generate new skills based on provided workflows.
4. Testing: Run "preview tests" to observe the agent’s chain of thought and verify its logic before full deployment.
5. Distribution: Share the agent with the team via a link or a company-wide directory, allowing others to "remix" or duplicate the agent for their specific needs.

4. Governance and Admin Controls

OpenAI emphasizes that builders and administrators maintain full control:

• Role-Based Access Control (RBAC): Admins define who can build, publish, and use specific agents.
• Compliance: All agent activity is captured in a centralized audit log (trace), which is exportable via API for monitoring.
• Security: Agents operate within the governance frameworks established by the organization’s IT and security policies.

5. Notable Quotes

• "Workspace agents are for teams. They're built for shared work for tasks that run in the cloud even when your computer is closed." — Christina
• "You don't need IT and engineering support to build out these workflows. We really want to make sure that these subject matter experts are the ones that can build out these flows." — Hojun

6. Synthesis and Conclusion

Workspace agents represent a shift from simple chatbots to autonomous, process-oriented AI workers. By combining Skills (the "how-to"), Tools (the "where"), and Memory (the "context"), these agents allow organizations to scale expertise and reduce manual, repetitive tasks.

Key Takeaways:

• Start Small: Use natural language to build and iterate.
• Leverage Memory: Enable persistent memory to allow agents to learn from previous interactions.
• Focus on Integration: Use Slack and Jira integrations to bring the AI directly into the flow of work.
• Research Preview: The feature is free until May 6, after which it will transition to a credit-based pricing model based on task complexity.