Same inbox audit task: Grok Bot vs Manus Cue
Running the same subscription-cancellation audit across years of email, Grok Bot delivered a usable first pass in under half the time, while Cue only improved after the author pointed out what it missed.
Today I ran the same task on @bot and on Cue from @ManusAI / @CueAgents. Task: go into my personal mailboxes, find every subscription, and tell me what I should cancel. Not “summarize my inbox.” Not “draft a nice table.” Audit a messy, multi-year trail of receipts, renewals, “your plan is about to renew,” and the emails I never opened because I already knew I was paying. Same mailboxes. Same prompt. No babysitting on the first pass. @bot finished in less than half the time. The first output was already usable: grouped by keep / review / cancel, with the evidence trail from the mail, not a vibe. Interaction was tight. Presentation was tight. I didn’t have to become the planner. Cue’s first pass was bad. I said the least agentic sentence in the English language: “You missed all of these.” Second pass got interesting. It recovered a couple that @bot didn’t catch. Credit where it’s due. Recall went up after I did the work of pointing at the holes. Overall? Not close. Speed, first-pass quality, how it asked for access, how it presented the kill list. @bot won the job. The technical read, because this is the part that actually matters if you build agents: Inbox audit is not a chat problem. It’s retrieval + entity resolution + time. You have to pull receipts across years, collapse “Netflix / Netflix Inc / http://Netflix.com” into one billing identity, distinguish trial, annual, family, and the thing you already cancelled but still gets a receipt, then rank by waste instead of by how loudly the email was written. Most agents fail on the first pass because they search like a junior analyst: latest folder, obvious brand names, the last 90 days. The expensive subscriptions hide in old threads, forwarded invoices, and “your card ending in…”. That’s why the second pass on Cue found extras. Once you hand it the misses, the search space collapses. The test is whether you needed that handoff. Latency is part of the product. An agent that is 2x slower on a personal finance task is not “more thorough.” It is making you wait while your own context window goes stale. And presentation is not cosmetics. A cancel list without evidence is a suggestion. A cancel list with the email, the amount, and the renewal date is a decision. Personal note, because of course there is one. I already pay too much for subscriptions. That was the whole point of the experiment. I now have a cleaner kill list than I had this morning. The joke writes itself: I used two paid agents to find out which paid agents and SaaS I should stop paying for. Last night @ManusAI 2.0 burned $15 of my own Anthropic tokens organizing a few Google Drive folders. Today Cue needed a correction pass on my mailbox. I am still a fan. I am also not easy to impress. If your agent cannot find the subscriptions quietly draining the card, it is not autonomous. It is a very expensive search box with manners. Now I have to go cancel something. Including, possibly, my dignity.