Find and book a business lunch restaurant
August 8, 2025 2025-08-08. Task: August 8, 2025 2025-08-08 Share Email article Share on LinkedIn Share on Twitter OpenAI recently released its agent mode for ChatGPT. Agent mode enables ChatGPT to navigate and manipulate websites designed for human use. This article documents my first impressions of this new interaction paradigm. While this tool represents a significant step towards mainstream AI-agent use, my simple evaluation revealed that agentic AI still has a long way to go. Note: On August 7 2025, OpenAI released GPT-5, the successor to the company’s collection of previous models. OpenAI’s Agent Mode was not updated alongside GPT-5, as they are separately developed AI models. In This Article: The Task: Book a Business Lunch Step 1: Search for Restaurants Step 2: Access the Restaurant Website Step 3: Clarify the Guest Count Step 4: Book the Reservation Step 5: Enter Details with Human Intervention Step 6: Submit Reservation Step 7: Confirm Reservation Overall: It Worked, But Beware the Caveats The Task: Book a Business Lunch Watch a recording of this interaction (about 5 minutes). This is a ChatGPT replay of this agent interaction; it's sped up and does not depict the actual duration I experienced. I chose the everyday task of booking a restaurant reservation. This was my starting prompt to ChatGPT: Find a restaurant suitable for a business lunch near 3100 Travis St, Houston, TX 77006 for next friday at noon. Is this a high-quality prompt? No. I kept it terse and deliberately omitted relevant details to simulate real-life use of everyday people who don’t fancy themselves as “prompt engineers.” (Writing highly detailed prompts poses a high interaction cost on users. In some of our recent studies of the prompting strategies of consumers, we’ve noticed that even experienced genAI users rarely include sufficient context and criteria in their first prompt.) Remember that the computer should adapt to the needs of the human, and not the other way around. I wanted to evaluate what this agent could perform with minimal handholding from a rushed and distracted user. This realistic-yet-weak prompt should have encouraged ChatGPT to request additional details before beginning. For example: What does “suitable” mean here? Are you going to talk with a prospective job applicant and want a quiet venue? Are you trying to impress a client's executive staff with something fancy? What’s the budget for the meal? Is this a casual lunch with a peer, or an expensive celebration? How will people be arriving? What’s more important: ample parking or access to mass transit? What type of food? Do attendees have preferred cuisines or dietary restrictions? How many attendees? Is this a 1-on-1 meeting, or will you need to rent out a private room in the restaurant? Instead of asking any of these clarifying questions, ChatGPT jumped immediately into the task with limited context. It focused on the necessary geographic proximity and operating hours but didn’t ask followups on relevant topics that would have significantly influenced its subsequent searches and suggestions. This behavior struck me as unusual, since ChatGPT’s “deep research” mode always asks clarifying questions before investing 8–15 minutes researching and writing a report. Step 1: Search for Restaurants ChatGPT spent 6 minutes conducting research on various restaurants by performing 10 web searches utilizing 96 sources — far more than any human could reasonably use to find a lunch spot for a business meeting. (Though, arguably, a human might not need to consult 96 different sources to make a well-informed decision.) ChatGPT’s search used mostly sensible sources for the task: Yelp, OpenTable, Instagram, and various restaurant-focused sites, and it took screenshots to document the process. ChatGPT did mention encountering some difficulties using these sites, such as dynamic features, anchor links that interfered with scrolling, and even the dreaded popup-upon-arrival. Excerpts of ChatGPT’s documented “thinking” during this process: The open table [sic] might be inaccessible due to dynamic features or cross-late issues. I'll use the computer tool to open it and bypass these limitations. It seems that scrolling has not progressed as expected, likely due to an anchor. A pop-up about Oporto is showing, and I need to close it by clicking the 'x'. I'll use the multi-action function to click the cross and close the pop-up window.; tools: ChatGPT Agent. Human role: The original post does not say how a person stayed involved.; limitations: Agent mode is significantly slower than human execution for simple web tasks and struggles with dynamic website elements like popups and complex selectors. It requires human intervention for personal data and final confirmation, reducing autonomy. Reliability is low for time-sensitive bookings due to processing delays and potential timeouts..
Tools
Published on Agent Work Atlas:
Overview
August 8, 2025 2025-08-08
Workflow
August 8, 2025 2025-08-08 Share Email article Share on LinkedIn Share on Twitter OpenAI recently released its agent mode for ChatGPT. Agent mode enables ChatGPT to navigate and manipulate websites designed for human use. This article documents my first impressions of this new interaction paradigm. While this tool represents a significant step towards mainstream AI-agent use, my simple evaluation revealed that agentic AI still has a long way to go. Note: On August 7 2025, OpenAI released GPT-5, the successor to the company’s collection of previous models. OpenAI’s Agent Mode was not updated alongside GPT-5, as they are separately developed AI models. In This Article: The Task: Book a Business Lunch Step 1: Search for Restaurants Step 2: Access the Restaurant Website Step 3: Clarify the Guest Count Step 4: Book the Reservation Step 5: Enter Details with Human Intervention Step 6: Submit Reservation Step 7: Confirm Reservation Overall: It Worked, But Beware the Caveats The Task: Book a Business Lunch Watch a recording of this interaction (about 5 minutes). This is a ChatGPT replay of this agent interaction; it's sped up and does not depict the actual duration I experienced. I chose the everyday task of booking a restaurant reservation. This was my starting prompt to ChatGPT: Find a restaurant suitable for a business lunch near 3100 Travis St, Houston, TX 77006 for next friday at noon. Is this a high-quality prompt? No. I kept it terse and deliberately omitted relevant details to simulate real-life use of everyday people who don’t fancy themselves as “prompt engineers.” (Writing highly detailed prompts poses a high interaction cost on users. In some of our recent studies of the prompting strategies of consumers, we’ve noticed that even experienced genAI users rarely include sufficient context and criteria in their first prompt.) Remember that the computer should adapt to the needs of the human, and not the other way around. I wanted to evaluate what this agent could perform with minimal handholding from a rushed and distracted user. This realistic-yet-weak prompt should have encouraged ChatGPT to request additional details before beginning. For example: What does “suitable” mean here? Are you going to talk with a prospective job applicant and want a quiet venue? Are you trying to impress a client's executive staff with something fancy? What’s the budget for the meal? Is this a casual lunch with a peer, or an expensive celebration? How will people be arriving? What’s more important: ample parking or access to mass transit? What type of food? Do attendees have preferred cuisines or dietary restrictions? How many attendees? Is this a 1-on-1 meeting, or will you need to rent out a private room in the restaurant? Instead of asking any of these clarifying questions, ChatGPT jumped immediately into the task with limited context. It focused on the necessary geographic proximity and operating hours but didn’t ask followups on relevant topics that would have significantly influenced its subsequent searches and suggestions. This behavior struck me as unusual, since ChatGPT’s “deep research” mode always asks clarifying questions before investing 8–15 minutes researching and writing a report. Step 1: Search for Restaurants ChatGPT spent 6 minutes conducting research on various restaurants by performing 10 web searches utilizing 96 sources — far more than any human could reasonably use to find a lunch spot for a business meeting. (Though, arguably, a human might not need to consult 96 different sources to make a well-informed decision.) ChatGPT’s search used mostly sensible sources for the task: Yelp, OpenTable, Instagram, and various restaurant-focused sites, and it took screenshots to document the process. ChatGPT did mention encountering some difficulties using these sites, such as dynamic features, anchor links that interfered with scrolling, and even the dreaded popup-upon-arrival. Excerpts of ChatGPT’s documented “thinking” during this process: The open table [sic] might be inaccessible due to dynamic features or cross-late issues. I'll use the computer tool to open it and bypass these limitations. It seems that scrolling has not progressed as expected, likely due to an anchor. A pop-up about Oporto is showing, and I need to close it by clicking the 'x'. I'll use the multi-action function to click the cross and close the pop-up window.
Reported result
The reservation was successfully booked, but the process was slow and inefficient. The agent took 6 minutes to search and 11 minutes to fill forms, compared to 2 minutes for a human. It failed to ask clarifying questions initially, leading to potential mismatches, and encountered technical delays with website interactions and timeouts.
Human role
The original post does not say how a person stayed involved.
Limitations
Agent mode is significantly slower than human execution for simple web tasks and struggles with dynamic website elements like popups and complex selectors. It requires human intervention for personal data and final confirmation, reducing autonomy. Reliability is low for time-sensitive bookings due to processing delays and potential timeouts.
Try it
Act as a business assistant using ChatGPT Agent to book a lunch. Input: [Location], [Date/Time], [Guest Count], [Dietary Restrictions]. Steps: 1. Search for suitable restaurants using Yelp/OpenTable. 2. Present top 3 options with menu/pricing details. 3. Wait for user selection. 4. Navigate to booking site. 5. Fill form, pausing for user to enter personal info. 6. Request explicit approval before submitting. Measure success by time taken vs manual booking and accuracy of dietary notes. If site blocks automation, stop and report error. Do not auto-submit without confirmation.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.