I Tested Every Major AI Agent for a Day. Here’s the Best One

A WIRED software reporter spent a single day putting six always-on AI agents through five basic tasks on an iPhone, scoring every attempt on a 1 to 5 scale to crown an overall winner and a separate 'most annoying' award. The agents were configured the night before the tests, and the experiments kicked off Thursday morning with a tech-news briefing task built around the Gear desk's weekly coverage meeting.

The exercise exists because the category has visibly shifted. A year ago, the same reporter sat through demo after demo of agents that promised to automate daily tedium and delivered nothing. Today the tools are racing to market and, more importantly, actually function — still error-prone, but markedly better at carrying a task through to completion and driving a browser. These agents differ from turn-based chatbots in that they nudge you first when they spot something useful, keep everything in a single thread that reads like a group text, and typically do their work inside a virtual browser acting as their own machine for internet access.

The lineup mixes free and paid entries across very different delivery shapes. Meta's Muse is free and app-based, as is Instinct, though Instinct requires an invite code from an existing user and operates over texts rather than an app. OpenAI's Dots sits inside ChatGPT behind the most expensive paywall tested. Google's Gemini Spark is a beta in the Gemini app; xAI's Grok Bot ships as a standalone app. Anthropic's Claude Duet is the newest arrival, launched as a wait-listed research preview the night before testing began.

OpenClaw was left off the list despite sparking much of this year's agent wave — deemed too clunky for ordinary users to set up — with smaller players like Hermes by Nous Research and Poke flagged for a possible future ranking. The reporter is explicit that this is not an endorsement: the software is nascent, and privacy and security risks are real, illustrated by an X user who linked agents to his bank account and work Slack, only to have one broadcast his balance to colleagues. Test exposure was deliberately limited to Google Calendar and Gmail.

First up was a test of which agent could assemble the strongest tech-news briefing from the Gear desk's Monday planning discussion and turn it into a compelling headline — the kind of morning-dossier use case these companies market hardest. What remains open: how the other four tasks shook out, whether a day-old research preview can beat established, higher-priced rivals, and whether invite codes, betas and paywalls matter more to early adopters than raw task performance as the category moves from demo to daily habit.