Grok vs. Dot vs. Muse vs. Claude Duet: which one crushes the competition?

The host says she has “personal agent fatigue,” but decided to run what she calls the first How I AI personal-agent benchmark. She asked Dot, Grok Bot, Muse, and Claude Duet to list the ten tasks they most commonly perform for her, then combined those suggestions and asked Claude to identify the most common, cross-cutting use cases. The goal was to test different capabilities rather than optimize for one particular agent. The nine selected tasks were drafting a Slack reply, cutting a short-form video, fixing a Sentry bug, creating a How I AI thumbnail, triaging inbox to-dos, making a family newspaper, explaining a physics problem to a fourth grader, shopping for shoes, and producing a ChatPRD launch video. She used one-line, human-style prompts rather than elaborate AI prompts. Codex used computer-use capabilities to operate Claude, Grok Bot, and Muse, while Claude computer use operated Dot because Dot was also needed as an agent under test. The host kept the process partly human-in-the-loop, nudging agents when computer use broke or they needed direction.

Before running the tasks, the host used Codex and browser-use tools to check that the agents had roughly the same connectors and skills. She describes this as an important practical lesson because her agents normally have inconsistent access to email accounts, business tools, and other services. However, she forgot to connect Muse to several business tools, including Slack and GitHub, and Muse therefore could not perform some tasks. She explicitly treats those failures as partly her own setup error rather than a pure test of the model. The benchmark was also not fully autonomous or blind. The host babysat the operator agents, dealt with broken computer-use actions, and judged the outputs live while recording the episode. She says Claude computer use was “hot garbage” and frustrating, but she needed it to operate Dot. At the initial scoreboard stage, Dot and Grok completed all nine tasks, while Muse and Claude completed seven. Muse was fastest on four of the nine tasks and generally finished quickly despite requiring more steering. Claude was the slowest overall and asked for permission four times, compared with once for each of the other agents, making its safety behavior notably more conservative.

For the Slack task, every agent with Slack access produced an acceptable reply agreeing to speak at an event for the host’s friend at Lovable. Dot stood out because it asked whether the event was remote and whether the host would be paid—details that matched her priorities. The host found Dot the most accurate, Grok the most in her voice, and Codex’s response the shortest but somewhat less accurate. Muse could not perform the task because Slack was not connected. The short-form video test exposed a major Claude failure: it would not download the raw video from Dropbox. Dot, Grok, and Muse all produced videos without the host touching them. Muse won on design, using a simple How I AI tag, the show’s caption layout, and punchier cuts. Grok added zoom effects but had unattractive overlays, while Dot looked reasonable but lacked the same punch. For the Figma thumbnail task, the result reversed: Muse made an ugly variation involving Blair peeking through curtains, while Grok generated fresh thumbnail concepts and headlines that the host liked. She says Grok is “100% the best” of these agents at using Figma. Muse also won the physics explanation task. All agents answered the velocity-over-time question correctly—answer C—and explained it adequately, but Muse responded in roughly half a second. The host says this speed has led her to move many simple “explain this” and lookup-style questions from ChatGPT to Muse.

Muse could not perform the coding task because it had no GitHub access. It also misunderstood a request to assign the issue in Linear to Devin, treating Devin as a person even though Devin was an agent. Dot, Claude Duet, and Grok each worked on the Sentry bug for about ten minutes. The host had Claude Opus 5.5 grade the fixes and concluded that Dot produced the closest match to the expected fix, while Claude found and fixed a second bug that the others had missed. Claude also recognized that Dot had already fixed the relevant issue and did not propose shipping unnecessary additional code. Grok performed worst because it loosened validation rather than properly solving the underlying problem. For inbox triage, all the agents found reasonable tasks, but the host cared about how the results were presented. Grok won by placing the to-dos in a newly introduced Grok document with checkboxes and a clean structure. Muse, which only had Gmail access, still found useful parenting tasks ahead of Halloween week. Claude put its suggestions into its own task-follow-up list, while Dot returned an unstructured message listing five things. The host therefore regarded Grok as the best personal chief-of-staff interface, even though Dot and Claude were stronger in coding itself.

The family-newspaper task produced sharply different results. Claude only had access to the host’s work calendar and responded with a Markdown outline rather than an actual newspaper, so the host considered it a clear failure. Dot generated what she called “trash” design. Grok produced a useful newspaper with a large “week at school” section, reminders, Bay Area sports, Halloween information, and weather. Muse impressed her most by creating a polished, family-oriented artifact with an appealing image of three boys, the family schedule on one page, and day-by-day weather and activity ideas on a second page. She says Muse could become excellent for this kind of task if tuned with a dedicated skill. For shopping, the host asked the agents to find affordable red Adidas Tokyo-style autumn shoes. Claude was blocked by most shopping sites, searched for white shoes with red stripes instead of red shoes with white stripes, and found the most expensive options. Dot barely attempted comparison shopping and did not explore alternate colors such as maroon. Grok was fast and found good options, matching the host’s prior experience using it for price comparisons. Muse also performed well but needed some steering. Her conclusion was that Grok and Muse are the useful shopping companions, while Claude failed and Dot did not make a serious attempt.

The final task was a launch video for ChatPRD’s Product Intelligence feature. Dot produced a presentation-like video that the host rejected as a failure. Muse lacked access to the product repository, invented much of the content, and added a British-accented voiceover; the host gave it a little credit for being ambitious but still judged it bad. Grok followed the host’s existing video skill and used the intended structure, but its layout contained errors and overflow. Claude Duet delivered the strongest result: it used the company branding, had no layout problems, clearly expressed the product’s value proposition, and used the ElevenLabs integration for a voiceover. The host says Claude “crushed” product walkthroughs, even though she would have preferred a female or recorded version of her own voice. The final ranking placed Grok Bot first, with a reported 93% high-quality completion score and wins in thumbnail creation, inbox to-dos, and shoe shopping. Muse came next for the personal and creative tasks it could access, particularly short-form content, the family newspaper, and fast homework explanations. Dot followed, though the host says its score was closer to Claude’s than to Muse’s because it struggled with creative work and was unpleasant to use. Claude Duet came last at 79%, despite winning the launch video and producing the most complete bug fix. Her practical recommendation is Grok for work and Muse for life, while Dot and Claude remain the preferred underlying tools for coding rather than the preferred general-purpose agents.

The host’s first general lesson is that consistent tool access is essential to any agent comparison. Muse’s missing Slack, GitHub, and other business connections directly limited its score, so an agent that appears weaker may simply have less access. She also strongly values speed: Muse’s rapid answers and fast completion of several tasks made it feel substantially better for everyday personal use, while Claude’s slowness caused her to forget what it was doing. She argues that an agent’s presentation layer matters as much as whether it discovers the right information. Grok’s documents, checkboxes, and organized artifacts were more useful than a plain message, and Muse’s attractive family newspaper showed the value of generating a finished artifact rather than an outline. Safety behavior is another differentiator. Claude asked for permission repeatedly and felt overly conservative, whereas the host likes Muse’s sidebar for tracking permission requests while still preferring a more proactive balance. Her final distinction is between an agent and the underlying coding tool or model. Dot and Claude remain the two strongest options for her coding workflows, but she would not necessarily use them as her general personal agents. She closes by saying she will continue using Grok Bot and Muse, while mostly ignoring Dot and Duet, and now has a repeatable framework for evaluating new agents.