Build Hour: Workspace agents in ChatGPT

Workspace Agents are presented as a significant evolution of ChatGPT’s capabilities, designed for team-based, complex, and long-running tasks. Unlike personal assistants, these agents operate in the cloud, integrating with multiple systems, files, and tools to automate entire workflows. They are powered by Codex and can be shared across an organization, improving consistency and efficiency.

Core Concepts and Architecture

Workspace Agents are distinct from other OpenAI offerings:

• Workspace Agents: Built for shared, collaborative work within teams. They run continuously in the cloud, independent of a user’s local machine.
• Codex: Suited for individual users who need a personal agent to accomplish tasks.
• Agents SDK: For developers who want to embed custom agent experiences directly into their own applications.

Key components of a Workspace Agent include:

• Tools and Connectors: Agents can be connected to various applications (e.g., Google Suite, Microsoft Suite, Slack, Jira). The builder has granular control over permissions, such as allowing read-only access to a calendar while granting write access to create documents or send emails.
• Skills: These are codified best practices, processes, or playbooks. Skills are reusable components that provide an agent with specific instructions on how to perform a task, ensuring consistency. You can import existing skills, create new ones using natural language, or use popular skills within your organization. An agent can leverage multiple skills.
• Memory: Agents feature a persistent file system that allows them to save notes, context, and outputs from previous runs. This enables them to learn and improve over time. Memory is scoped to the context; for example, each user has a private memory for their agent instance in ChatGPT, while an agent in a shared Slack channel has a shared memory for that channel.

Demo 1: Automated Meeting Preparation Agent

This demonstration showcased a personal productivity agent named “Auto” designed to automate the manual work of preparing for customer meetings.

Workflow:

1. Trigger: The agent is scheduled to run every morning.
2. Information Gathering: It connects to Google Calendar to identify the day’s meetings.
3. Research: For each meeting, the agent performs research by searching the web and scanning internal documents in Google Drive for relevant customer context, account notes, and company details.
4. Brief Creation: Using a predefined template (provided as a “skill”), the agent generates a structured meeting brief in a Google Doc for each meeting. The brief is formatted with headers, tables, and bullet points, including sections like an executive summary, customer snapshot, and meeting objectives.
5. Notification: The agent composes and sends an email via Gmail containing a summary of the day’s meetings, with direct links to each of the newly created briefs.

This allows the user to be fully prepared for their day without spending hours on manual prep work. The agent can be shared, and other team members can duplicate and “remix” it for their specific needs (e.g., swapping Google Drive for SharePoint).

Demo 2: Software Review Agent in Slack

This agent is designed for team-level service, managing high-volume, time-sensitive software requests within a company’s Slack channel. It acts as an intelligent gatekeeper and facilitator for the IT and procurement teams.

Workflow:

1. Trigger: An employee makes a request for a new software tool (e.g., “Screen Studio”) in a designated Slack channel.
2. Analysis and Research: The agent receives the message and begins its process. It researches the requested software’s capabilities and compares it against an approved software list stored in a Google Sheet.
3. Reasoning: The agent reasons across multiple data points, including:
   • The company’s procurement policies (codified in a “skill”).
   • The functionality of existing, approved tools.
   • The current utilization and license availability of comparable tools (e.g., it identifies that “Bloom” is an approved alternative but all licenses are currently in use).
4. Action and Response: Based on its analysis, the agent takes a specific action. In the demo, it determined that the request required human review because the preferred internal alternative had no available licenses.
   • It responds directly in the Slack thread, clearly stating its decision and rationale.
   • It provides the user with context on the approved alternative and the reason for the escalation.
   • Crucially, it automatically creates a task in Jira for the IT team to review the request, attaching all the relevant context.

This process saves the IT team significant time, reduces duplicate tool spending, and provides employees with a fast, transparent, and actionable response.

Building, Governance, and Administration

Building an agent is a conversational, no-code process within the ChatGPT interface. A user describes the desired workflow in natural language, and ChatGPT generates an agent plan, wires up the necessary tools and skills, and writes the instructions.

Key development and administrative features:

• Preview and Testing: Builders can perform test runs to observe the agent’s behavior and “chain of thought” in a detailed trace, allowing for iteration and refinement before deployment.
• Admin Controls: For Enterprise and Edu plans, administrators have role-based access controls (RBAC) to define who can build, publish, and share agents, and which applications are available for connection.
• Monitoring: All agent activity is centrally logged and auditable. The Compliance API allows admins to monitor and manage usage across the organization.
• GPTs to Agents: Workspace Agents are positioned as the next stage of Custom GPTs, evolving from simple templates to powerful, multi-step workflow automators. An automatic converter from GPTs to Workspace Agents is planned.

• Tags:
• ChatGPT
• Feature
• Automation
• Codex
• No-Code
• Workspace Agents