OpenAI Agent Mode: 58 Minutes for Cupcakes—Should You Trust It?

OpenAI’s new “agent mode” delivers real capability gains—especially for finance-adjacent workflows like building and filling Excel templates—but it still behaves like an overthinking, slow “intern” that requires frequent supervision. The cupcake example is the clearest warning sign: ordering a custom batch online took about 58 minutes and multiple authentication handoffs, even though the task was straightforward. For everyday users, that latency and babysitting burden undercut the promise of autonomous assistance.

The strongest practical case for agent mode is its fit with Excel-heavy work. There’s a long-standing gap between AI systems and spreadsheet workflows: models can often read spreadsheets, but reliably generating correct formulas and handling real-world complexity—especially large, multi-thousand-row “spreadsheet from hell” files—has been difficult. Agent mode aims at that blind spot by connecting to tools like Excel and Google Drive and performing the research and execution needed to populate templates with correct methodology, formulas, and numbers. Investment bankers, in particular, are described as lining up for this kind of background automation, because finance teams already live inside Excel.

Still, the transcript argues the release doesn’t solve the hardest agent problems—autonomy, discernment, and safe obstacle navigation. The design leans heavily on guardrails and user intervention, which is framed as a liability-management strategy for high-stakes actions (like purchases). That supervision-first approach creates a mismatch with how people actually want assistants to work: most users don’t want to stand over an intern’s shoulder, and they want quick, reliable help for routine tasks.

Security concerns also loom. Sam Altman is cited warning that agent mode could be hijacked via prompt injection delivered through email—an “email as a prompt injection attack” where text inside an email could manipulate the agent when it opens the message. The transcript extends the risk beyond email, noting that similar low-contrast or hidden instructions could be embedded on other websites, including tactics already used in research and hiring contexts to influence automated evaluators. In response, the transcript calls for agents that can recognize malicious inputs and reason through obstacles without needing constant oversight.

The broader critique is about tradeoffs and timelines. Agent mode is portrayed as part of a longer, decade-scale effort to build a general-purpose GUI-navigating agent—analogized to how Tesla builds cars to navigate streets. But collecting data “in the wild” effectively turns users into guinea pigs, and the value delivered so far may not match the cost for most people outside finance. The ideal assistant, the transcript suggests, is one that users touch daily: fast, accurate, and largely hands-off. Agent mode, by contrast, is slower, more interactive, and more constrained.

The takeaway is mixed: agent mode is a step forward and can be useful—particularly for Excel-centric finance tasks—but it’s unlikely to become broadly adopted soon because its current UX and safety model still require too much supervision, and because the autonomy and security gaps remain largely unresolved.