Reviewer tests three AI agents, finds two fail at daily tasks

By Splitfeed EditorialLast updated Oct 16, 2026, 14:20 UTC

On October 16, 2026, PCMag published a hands-on review in which a single reviewer allowed three AI agents to take over management of real daily-life tasks. Of the three agents tested, only one completed its assigned work without failing. The other two fell short in ways significant enough for the reviewer to call them outright failures. The piece ran across PCMag's US, UK, and Australian editions simultaneously, giving it a wide consumer-facing audience.

The test matters because AI agents, meaning software that can autonomously plan and execute multi-step tasks on a user's behalf, are increasingly being marketed as ready for everyday use. A two-out-of-three failure rate on ordinary daily tasks is a concrete data point against that claim, at least as measured by one reviewer's methodology. What remains genuinely in dispute is whether the failures reflect fundamental limitations of the technology, problems specific to how those two products were implemented, or simply a mismatch between the reviewer's chosen tasks and what each agent was designed to handle. Without the underlying detail of which agents were tested and what the tasks were, it is difficult to generalise the result, but the headline finding is stark enough to fuel the broader debate about whether AI agents are ready for mainstream adoption.

Because no camp has published reactions to this story yet, what each would typically argue can only be anticipated. Pro-AI voices would likely frame the one success as proof of concept and argue that a rapidly improving field should not be judged by the weakest products in any given test. Anti-AI voices would likely treat the two failures as evidence that industry hype has outrun actual capability, pointing to real-world consequences when people delegate important daily tasks to unreliable systems. A middle-ground position would probably call for standardised, transparent benchmarks for agent evaluation rather than individual reviewer tests, arguing that neither blanket enthusiasm nor blanket skepticism is warranted from a single hands-on article.

The argument would sharpen considerably if the agents named in the full review release updated versions responding to the criticism, or if independent benchmarks covering agentic task completion are published in the months following. A follow-up test using the same methodology on newer model versions would offer the clearest like-for-like comparison.