Search for cheap round-trip flights from JFK
ChatGPT agent returned a table of dates and prices per destination that matched on checking, but gave no booking links so the reporter had to search manually.

ChatGPT agent returned a table of dates and prices per destination that matched on checking, but gave no booking links so the reporter had to search manually.
A tester had ChatGPT Agent build a generative AI careers deck; the result was basic with hard-to-read tables, and the Canva connector was read-only so it couldn't improve it.
A RevOps practitioner built a Zapier Agent Slackbot to answer reps' questions; it took an afternoon to work and he concluded agents are imperfect.
The writer used ChatGPT Agent to search Gmail for renewal emails, extract service, date and cost into a list, and ask before unsubscribing; it also found cheap gas nearby.
A CNET reporter had ChatGPT's agent research stylists and request booking details, but stopped it before it booked.
A commenter had ChatGPT build a spreadsheet for her genealogy research, and agent mode was recommended for compiling and formatting findings into a usable document.
In Agent mode the author had ChatGPT search local plant shops and fill the checkout form, stopping for him to click buy, then reported the order with tracking.
An SE Ranking marketer tested ChatGPT Agent on marketing tasks including a gated HubSpot report; it fully succeeded in only two of five runs.
A writer had the agent cancel Peacock, Disney and Netflix after logging in; it needed a nudge and may be blocked on some sites.
Author used ChatGPT Agent to reorder an item from a photo of its wrapper, handling the whole Amazon purchase until Amazon blocked agents.
A ChatGPT workflow finds bills in statements, looks up competitor prices, emails providers via Gmail asking them to match, and summarizes outcomes, with user review before sending.
Author had ChatGPT Agent pick and book a Gatwick hotel on Booking.com; it reached the card confirmation screen, then needed the author to take over.
ChatGPT Agent worked 50 minutes and reported adding five lamps to the cart, but the user's own cart was empty as it ran on a separate computer.
ChatGPT Agent searched academic databases for recent open-access empirical papers, grouped them by field, wrote an annotated bibliography, noted research gaps and assembled a draft.
A reviewer ran ChatGPT Agent Mode through five business tests, including equity portfolio analysis, marketing attribution and real estate scenarios, noting where it needed babysitting.
Testers had ChatGPT Agent navigate the WHO site, extract the latest data and organise it by indicator, country and trend; it often fails on logins or dynamic pages.
A reporter had ChatGPT Agent plan a 4-day San Diego trip: it searched flights, compared hotels, built an itinerary and pre-filled booking forms, stopping before purchase.
A reviewer's dot flagged a failed payment unprompted but only read and reported it; a shopping site blocked it and follow-through was unfinished.