Cached at:
06/08/26, 07:44 PM
### TL;DR
OpenAI Solutions Engineer Lee Spacagna demonstrates how to quickly build a "Chief of Staff" AI Agent using the new ChatGPT Workspace Agent, automating the data, tools, and workflows in a team’s daily operations — filling the mid-level automation gap between personal chatbots and large-scale system changes.
---
## Two Paths for AI in Finance — and the Missing Middle Layer
Every day when working with financial institutions, customers ask: where can AI really change my business operations? Today there are two paths for AI adoption:
- **Bottom-up**: Give employees ChatGPT and Codex tools to use AI in their daily work.
- **Top-down**: Build new products at the company system level, such as improving customer service or optimizing operational support.
But there's a missing middle layer — **team and department-level automation**. That's exactly the gap the new ChatGPT Workspace Agent aims to fill.
---
## What Is an Agent?
An Agent is an AI system you can delegate meaningful tasks to, not just ask questions. Using existing tools (email, calendar, productivity apps, etc.), an Agent can get work done like a human. Over the past few months, Agent capabilities have jumped dramatically — especially after the GPT 5.5 release. Today, Agents can handle complex tasks that used to take hours or even days, and they can execute them from start to finish.
With Workspace Agent, users can elevate custom GPTs to a much more powerful level: the new Agent builder integrates shared apps, skills, and deployment into one platform, allowing Agents to operate in the same place where work happens.
---
## Building a "Chief of Staff" Agent (Demo)
### 1. Use a Template and Connect Tools
On the left side of the standard interface, there's a new "Agent" option. Select the existing "Chief of Staff" Agent template — it already includes a set of instructions, tools, and capabilities. Then connect Microsoft tools: Outlook Calendar, Teams, and Outlook Email.
**Key highlight**: The instructions were automatically written by another Agent — no prompt engineering or technical skills required. Business users can use natural language to have one Agent build another Agent for themselves.
### 2. Customize the Daily Briefing
The initial version is already usable, but the user wants the Agent to run every morning at 9 AM, check meetings, apps, and overnight emails, and generate a daily briefing. Simply give the instruction in natural language: "I need it to run at 9 AM." Within seconds, the Agent is customized.
### 3. Test and Deploy to Teams
Two starter prompts (built-in capabilities) appear below. Select "Prepare today's briefing," asking the Agent to use all available information to produce a concise briefing that highlights priorities, decisions, blockers, and next steps, then post it to the "daily prep" channel of the CFO team's workspace.
The Agent starts — it fetches emails, calendar, and other resources, cross-referencing the day's meetings. On first run, it requests authorization to post to Teams. After granting access, the briefing is successfully posted to Teams' daily prep channel.
In just a few minutes, a Chief of Staff Agent — connected to daily tools, customized with instructions, and serving the entire team — is built.
### 4. Add Meeting Preparation Capability
The team is exhausted from back-to-back meetings and has no time to prepare beforehand. The user wants the Agent to proactively research before each meeting: who is attending, what's the latest status, and what are the meeting objectives.
First, add more tools: connect from SharePoint (where company information and notes are stored) and Salesforce (CRM and customer context). Once done, these apps are now wired into the Agent.
Then add "skills." Skills are a way to capture snippets of information and instructions for performing key tasks — essentially turning team tacit knowledge and routines into repeatable workflows. Two skills already exist ("Chief of Staff" and "Final Briefing Formatting"), so add an existing "Meeting Preparation" skill, which defines the information structure, key details, sources, and where to post.
### 5. Update Instructions and Deploy
Using natural language, tell the Agent next to you: "I just added Salesforce and SharePoint, I need a new capability — make the Agent generate a quick meeting brief, and only provide information for the *next* meeting." The Agent updates accordingly and automatically adds a third starter prompt below: "Prepare my next meeting."
When the Agent runs, it extracts context from sources like Salesforce and SharePoint, then combines everything into a concise briefing following the skill's format.
---
## Daily Use of the Agent and Personal Benefits
Lee shares his own experience: every morning he runs an Agent to check all overnight emails, scope updates, things he promised yesterday on Slack or over the phone, and context from call logs. By the time he arrives at the office, reply drafts for all emails are ready — just click "send" or "approve" to send to clients. This saves him the first hour of every day, completely transforming how he works.
For the team, where before there were never enough people to prepare for meetings, now everyone can attend meetings as if they had their own Chief of Staff.
---
## Managing Agents at Scale: The Frontier Platform
What about managing hundreds or thousands of Agents? **Frontier** is OpenAI's platform for deploying and managing Agents at scale. It connects typically siloed data warehouses, CRMs, and internal applications, giving AI colleagues the same shared context as the team. Agents reason over data, run code, use tools, and take action — all within a governed environment.
Key feature: The system learns from interactions and evaluates its own performance over time — the more it does, the better it gets, just like an employee in an enterprise.