Agent Use Cases › Comparisons

Test AI agents on your own recurring jobs before you pick one

Claire Vo, host of the How I AI podcast, put four personal agents (Grok Bot, ChatGPT’s Dots, Meta’s Muse and Claude Duet) through nine jobs she actually needed done. No agent won everything: Grok Bot took inbox to-dos, thumbnails and shopping, Muse the family newspaper and short video clip, Dots the Slack reply, and Duet the product launch video.

Done with: Grok Bot, ChatGPT Dots, Meta Muse and Claude Duet, with Claude computer use running the tests and Codex checking each agent’s connections

How did Claire run the comparison?

Claire, founder of ChatPRD, asked each agent for the top 10 tasks it does for her, then had Claude build a benchmark around the most common ones. The nine jobs: a Slack reply, inbox triage, a code bug fix, cutting a short video clip, YouTube thumbnails in Figma, a printable family newspaper, a fourth-grade physics question, shopping for red Adidas Tokyos and a product launch video. She wrote light prompts the way she’d really ask (“Keep it in my voice. Do exactly what I would do”) plus answers to likely follow-up questions and nudges for a stalled agent. She used agents to drive the agents in parallel, and had Codex check each agent’s connections and skills. That check mattered: Muse didn’t have Slack or GitHub connected and Claude only had her work calendar, so some “failures” were really missing access. She then judged the finished work. Grok Bot’s inbox to-dos came as a document with check marks, which she found more usable than a list in a chat message. Muse made the most appealing family newspaper and the best-paced clip, and answered the physics question in about half a second. Dots was the only agent to ask whether a speaking invitation was paid and remote. Duet wouldn’t download the footage for the short clip but made the best branded launch video. On the code fix, Dots came closest to the intended fix and Duet found a second bug, while Grok Bot’s fast fix loosened a validation check, which she flagged as the concerning one.

“Different agents belong in different parts of life.” Her split: Grok Bot for work, Muse for personal tasks.

What did the test show?

How can you try it yourself?

1. List the 5 to 10 jobs you actually repeat each week. Asking each agent what it already does for you is a quick way to find them.
2. Write each prompt the way you’d really ask, and decide up front what a usable result looks like: a checklist, a draft reply, a file or a product with the right color and price.
3. Before you start, check that every agent has the same apps connected (email, calendar, Slack, files). Otherwise you’re comparing access, not ability.
4. Run the same prompt in each agent, and note how long it took, how often it asked for approval and what you had to fix.
5. Give each job to the agent that did it best, keep human review on anything that ships, spends or changes code, and rerun the test when the tools update.

I'm comparing AI agents on my real work. Here are the jobs I repeat each week: [list]. For each job, write a one-line prompt in my voice, describe what a usable finished result looks like (a checklist, a draft, a file), and list the apps or accounts an agent needs connected to do it. Then make me a score sheet with columns for: finished or not, time taken, number of approval requests, mistakes I had to fix, and whether I'd use the result as it is.

← All Agent Use Cases