Find a vintage lamp on Etsy
OpenAI used DevDay on September 29, 2026 to move agents from a demo you watch into a service that runs in the background: Dots, always-on assistants living inside ChatGPT, each with its own cloud computer. Task: The case describes a suggested implementation where a user tasks ChatGPT Agent (specifically the earlier version or Dots) to search for a specific product, such as a vintage lamp, on Etsy. The agent uses its own cloud computer and browser to perform multi-step research, filtering, and comparison. The human role involves defining the goal, reviewing the agent's reported progress, and manually verifying the actual state of the shopping cart or website, as the agent operates in an isolated environment without the user's login credentials.; tools: ChatGPT Agent. Human role: The original post does not say how a person stayed involved.; limitations: Agents run on isolated cloud computers and cannot access the user's browser logins, cookies, or payment methods. Financial actions, account management, and irreversible transactions are restricted or blocked by safeguards like Watch Mode. Users must manually verify any claimed actions (like adding to cart) because the agent's environment is separate from the user's. Source reliability is secondary, based on a specific hands-on test rather than broad statistical data..
Tools
Published on Agent Work Atlas:
Overview
OpenAI used DevDay on September 29, 2026 to move agents from a demo you watch into a service that runs in the background: Dots, always-on assistants living inside ChatGPT, each with its own cloud computer.
Workflow
The case describes a suggested implementation where a user tasks ChatGPT Agent (specifically the earlier version or Dots) to search for a specific product, such as a vintage lamp, on Etsy. The agent uses its own cloud computer and browser to perform multi-step research, filtering, and comparison. The human role involves defining the goal, reviewing the agent's reported progress, and manually verifying the actual state of the shopping cart or website, as the agent operates in an isolated environment without the user's login credentials.
Reported result
The source reports a failure in execution fidelity: the agent claimed to have added five lamps worth $825 to the cart, but the user's actual cart remained empty because the agent lacked login access and ran on a separate virtual computer. The agent subsequently refused to place the order due to restricted payment access. No successful purchase or verified outcome is reported; the result was a false positive regarding task completion.
Human role
The original post does not say how a person stayed involved.
Limitations
Agents run on isolated cloud computers and cannot access the user's browser logins, cookies, or payment methods. Financial actions, account management, and irreversible transactions are restricted or blocked by safeguards like Watch Mode. Users must manually verify any claimed actions (like adding to cart) because the agent's environment is separate from the user's. Source reliability is secondary, based on a specific hands-on test rather than broad statistical data.
Try it
I want to test the capability of ChatGPT Agent to perform product research on [replaceable input: e.g., Etsy, Amazon] for [replaceable input: e.g., vintage lamps, ergonomic chairs]. 1. Define the search criteria: [replaceable input: e.g., price under $200, condition 'used', specific style]. 2. Instruct the agent to search, filter, and summarize the top 5 options with links and prices. 3. Ask the agent to simulate adding these items to a cart and report the total cost. 4. CRITICAL: Do not allow the agent to attempt checkout or payment. 5. After the agent reports completion, I will manually open the links in my own browser to verify if the items are actually available and if the agent's summary matches reality. 6. Output format: A table with columns for Item Name, Price, Link, and Agent's Claimed Status. 7. Success metric: The agent provides accurate links and prices. Failure metric: The agent claims actions it cannot perform (like logging in) or provides hallucinated availability. 8. If the agent fails to find items or hallucinates, stop and report the error. Do not proceed to any purchase steps.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.