Build Hour: Workspace agents in ChatGPT

**TL;DR:** OpenAI demonstrated how to use natural language to build Workspace Agents in ChatGPT during Build Hours, using a meeting preparation assistant as an example, showing the full workflow of no‑code creation, tool configuration, preview runs, and team sharing. ## Build Hours Overview Build Hours is OpenAI's live stream series designed to help teams get more value from OpenAI products through real‑world examples and practical tips. This episode focuses specifically on Workspace Agents, available (as a research preview) in ChatGPT Business, Enterprise, Edu, and ChatGPT for Teachers plans. ## What Are Workspace Agents? Christina explains that Workspace Agents are agents powered by Codex, designed to handle complex, long‑running work that spans multiple systems. They can access files, code, and tools, run continuously in the background, and keep working even after you close your laptop. Agents have memory, can be guided through conversation, and improve over time. Getting started is simple: click Agents in the ChatGPT sidebar, describe a workflow your team is already doing, and ChatGPT helps turn it into an Agent. ### Differences from Codex and Agents SDK - **Workspace Agents:** Designed for teams to share work. Tasks run in the cloud and continue even when devices are off. - **Codex:** Suited for individuals working with their own personal agents. - **Agents SDK:** For teams that want to integrate custom agents into their own products and customer experiences. ## OpenAI Internal Use Cases Several teams are already using Workspace Agents: - **Marketing team:** Converts product briefs directly into websites by extracting requirements from Google Docs and code. - **Accounting team:** Assists with month‑end closing, making it faster and more consistent. - **Finance team:** Conducts vendor risk reviews by researching vendors and generating structured reports. ## Demo 1: Meeting Preparation Assistant Hojun walks through building a meeting preparation agent (named Auto) that checks the calendar each morning, researches client information on Google Drive or the web, and then generates a meeting briefing and sends an email. ### Build Process 1. **Start with natural language:** Enter a prompt describing the use case (e.g., “Help me prepare for sales meetings, list the tools I need, I have a template, and finally email me a summary of upcoming meetings with a link to the full briefing”). 2. **ChatGPT generates the Agent plan:** It first creates an outline for the user to review, then automatically configures schedules, apps, tools, and instruction sets. 3. **No‑code interaction:** Users can give feedback directly in chat, and ChatGPT adjusts the configuration without needing IT or engineering support. 4. **Configure tool permissions:** In the example, the Agent only needs to reference Google Calendar – Hojun disables write access to ensure the Agent cannot modify events. Similarly, the scope of tools like Gmail can be controlled. 5. **Add skills:** Skills can come from other tools, existing processes on the platform, or be generated by ChatGPT. Hojun demonstrates having ChatGPT generate a skill and add it to the Agent’s playbook for formatting meeting briefs (tables, headings, bullet points). ### Preview Run - When starting a preview test, users can see the Agent’s chain of thought and every step in real time (e.g., accessing the calendar, searching Drive, generating a document). - Preview runs don’t have to be shown to end users, but they are very helpful for builders to debug and verify the workflow. ### Result Hojun shows an email from Agent Auto containing a beautifully formatted meeting briefing, including: - Executive summary - Client snapshot - Meeting objectives (from the template) The Agent automatically completes its daily tasks in the background – the user only needs to check in when free, saving significant manual preparation time. ### Schedule and Sharing After building, you can: - **Schedule execution:** Set it to run daily at a specific time, or trigger on demand. - **Share with the team:** Let other members customize and use it according to their own workflows. ## What Follows (Unfinished) The transcript ends here and does not include Christina’s getting‑started tips, the differences between Agent and GPT, admin controls, or the second demo of a software approval assistant. Based on the original plan, these topics were supposed to be covered after the demo, but this article faithfully captures only the content that was presented based on the available transcript. --- **Source:** https://www.youtube.com/watch?v=kktBVmjA19A