I tested 4 AI agents on 9 real tasks. Here’s who won.

Summary

The How I AI host benchmarked four personal AI agents—Grok Bot, Muse, Dot, and Claude Duet—on nine everyday and work-related tasks. The goal was to compare their practical usefulness, speed, tool use, and output quality, rather than test them with lengthy, idealized prompts.

How the benchmark was run

• Each agent received short, similar prompts, often asking it to act “in my voice” or do what the host would do.
• The host used computer-control tools to operate the agents: Claude ran Dot, while Codex ran Grok Bot, Muse, and Claude Duet.
• The host tried to provide comparable tool access but missed some connections. Most notably, Muse lacked access to Slack and business systems such as GitHub.
• The test was not fully autonomous or blind. The host intervened when computer use stalled or crashed and noted that differences in memory, permissions, and tool access affected the results.

Results by task

1. Draft a Slack reply: Dot gave the most personalized response, asking about payment and whether the event could be held remotely. Grok best matched the host’s voice. Muse could not complete the task because it lacked Slack access.
2. Edit a short video from Dropbox footage: Muse produced the host’s favorite edit, and Dot also did well. Grok used zooms but had poor overlays and took longer. Claude Duet failed to upload and edit the video.
3. Fix a Sentry bug: Dot made the closest direct fix. Claude found and fixed an additional issue and recognized that Dot’s change did not need to be duplicated. Grok weakened a check rather than fixing the underlying problem. Muse lacked GitHub access.
4. Create a Figma thumbnail: Grok won by proposing fresh thumbnail and title ideas. Dot and Claude mostly adapted existing templates, while Muse’s result was poorly received. The host particularly praised Grok’s Figma performance.
5. Triage email into to-dos: Grok presented tasks in the clearest, most usable checklist format. Muse identified relevant family tasks from Gmail, and Claude offered task tracking. Dot’s plain list was the weakest presentation.
6. Make a family newspaper: Muse produced the most appealing and creative result, including an illustrated layout. Grok made a useful, informative newspaper. Dot’s design was poor, while Claude produced a Markdown plan rather than the finished newspaper, partly because it lacked access to the family calendar.
7. Explain a physics problem to a fourth grader: All agents answered correctly. Muse was exceptionally fast and was the host’s favorite for this kind of quick explanation.
8. Shop for shoes: Grok was strongest at finding and comparing options. Muse also did well, with some guidance. Claude was frequently blocked and found unsuitable, expensive alternatives; Dot did little comparison.
9. Create a product-launch video: Claude Duet was the clear winner, accurately communicating the product’s value proposition and using the brand well. It also used an ElevenLabs MCP integration for narration. Grok followed the intended skill but had layout problems. Dot’s video was poor, and Muse lacked access to the project repository and invented unsupported details.

Overall takeaways

• Grok Bot was the overall benchmark winner, reported as completing 93% of tasks at high quality. It won the thumbnail, email-triage, and shoe-shopping tasks and performed well across a broad range of work.
• Muse stood out for personal and creative tasks and was the fastest in four of the nine tasks. The host highlighted its video editing, family newspaper, and homework explanations, while noting that missing work-tool access limited its results.
• Claude Duet was the slowest overall and asked for permissions more often than the other agents. It struggled with several practical tasks but excelled at the bug fix and product-launch video. Its final reported score was 79%.
• Dot showed strengths in personalized writing and coding but disappointed on several creative and organizational tasks.
• The host’s practical recommendation was Grok Bot for work and Muse for everyday personal use. For coding, the host considered the underlying OpenAI/Codex and Anthropic/Claude capabilities strong, though not necessarily a reason to choose those agents for every workflow.

The host stressed that tool access, presentation, speed, and permission handling all materially affect the agent experience. The benchmark is best treated as a hands-on comparison, not a controlled scientific test, because access was not perfectly equal and the host and operator intervened during runs.

Main speakers and sources

• Main speaker: The How I AI host, who designed and narrated the benchmark.
• Agents tested: Grok Bot, Muse, Dot, and Claude Duet.
• Computer-use operators: Claude ran Dot; Codex ran Grok Bot, Muse, and Claude Duet. Codex also handled timing and result collection.
• Evaluation: The host reviewed the task outcomes live. Claude was also used to assess code fixes and compile the final ratings.