Four Personal AI Agents on Nine Tasks: A User’s Test

Claire Vo compares four personal agents—Grok Bot, Dot, Muse and Claude Duet—on nine tasks in October 2026. This is her personal, exploratory test, not a standardized or independently replicated benchmark.

She asked the agents for common tasks and selected overlapping use cases. The prompts were short and written in her usual style, covering work, creative production and family needs.

An operator agent controlled each assistant through computer use. Claude operated Dot, while Codex operated the other three; Vo also intervened when automation broke or needed nudging. Different operators are a significant confounder.

Vo tried to align available connections and skills but forgot to connect Muse to several business tools. Assistants also had different histories and memories, and she scored the outputs knowing which system produced them.

The nine tasks include a Slack reply, short video editing, a Sentry bug fix, a Figma thumbnail, inbox triage, a family newspaper, a child’s physics question, shoe shopping and a product launch video.

Vo reports that Dot and Grok completed all nine tasks, while Muse and Duet completed seven in her setup. Muse lacked some connections, and Vo says the automation itself sometimes needed help; completion counts should be read in that context.

She reports that Muse was fastest on several tasks and Duet was often slower, while permission requests varied. Speed and asking for approval are different dimensions from correctness and appropriateness.

For the Slack reply, Muse lacked access. Vo says the available drafts were broadly usable; she preferred Dot’s inclusion of compensation and remote participation because those matched her personal preferences.

For a short-form video, Duet did not complete the requested download and edit in this run. Vo preferred Muse’s cuts and layout, placed Dot next, and disliked aspects of Grok’s overlay. Her visual ranking is subjective.

On a Sentry bug, Muse had no GitHub connection. Vo says Dot came closest to the expected fix, while Duet also identified another bug; she criticizes Grok’s fix for loosening validation. These are assessments of one issue, not a general coding benchmark.

For a Figma thumbnail, she disliked Muse’s design, found Dot and Duet usable but conservative, and preferred Grok’s alternative ideas. The result depends on her template, materials and aesthetic preferences.

All the tools with relevant access found inbox tasks, but Vo preferred Grok’s document format. Muse saw only her Gmail, while she found Dot’s brief message less useful for this workflow.

For a family newspaper, Vo liked Muse’s visual presentation and Grok’s organization. She says Duet had only her work calendar and produced an outline, while Dot’s layout did not meet her taste. Unequal access affects this comparison.

All four answered the selected velocity-versus-time homework question correctly according to Vo. She preferred Muse because it responded quickly; one exercise does not measure overall educational reliability.

For shoe shopping, she preferred Grok’s comparison and found Muse nearly as useful with some steering. Duet encountered blocked sites and mismatched the requested colors; she judged Dot’s search incomplete.

For a product launch video, Duet produced the version Vo preferred. Grok followed more of her established style but had layout overflow; Muse lacked repository access and invented details, demonstrating why output claims need checking.

Vo feeds her spoken judgments into Claude for a final ranking, where Grok comes first and Muse follows. She says she does not know exactly how the displayed percentages were calculated; they should not be treated as reproducible accuracy rates.

Her practical takeaway is to choose tools by task and to check connectors, speed, output presentation and approval behavior. She personally favors Grok for several work tasks and Muse for family-oriented tasks in this setup.

For coding, Vo says the Dot and Duet results were strongest in her test, but she would generally use the underlying coding tools directly. All conclusions remain specific to the prompts, access, operators and judgments used here.