Hands-on with personal agents: simple life admin still unreliable
The author has tried nearly every personal agent and finds models still can't reliably handle price checks or email replies without heavy guidance, though progress in eight months is huge; personalization and memory remain the shared gap.
all inspired by @openclaw back in December - February! FWIW I dont agree with @signulll that the intelligence level is there yet for normal workflows. I've tried almost all of the personal agents and none of the models are (IMO) smart enough to do fairly simple life admin tasks (checking prices, replying to emails) without needing lots of guidance about the user's intent, running into technical errors, or just lacking common sense. However, I think we'll see models becoming smart enough to *reliably* take on your daily tasks within the next 6 months. To see how fast we are progressing: back in February I tried to get my OpenClaw to order groceries for me on Picnic (the first OpenClaw add-on? @MaxSumrall can confirm!) and it was a complete disaster, I had to stay up until 4am babysitting it and nothing worked, and despite being told "check what I usually buy and buy the same", it bought complete different stuff. Today, this is *basically* a oneshot with Opus/Fable if you prompt it carefully (although I've tested this and even Fable is lazy and doesn't properly check your prior purchases unless you yell at it). Huge progress in 8 months. The same will hopefully happen for most of the tasks agents struggle with today. One thing that I think *all* of the products miss, however, is personalization and truly knowing the user. Currently everyone does memory and sort of some text based prompts; which doesn't work because today's models write bad memories, riddled with errors. I'm fixing this with @OpenTrawl (opensource - MIT) by just letting the agents search your messages on your Mac - the "Bitter Lesson-pilled" solution. I'm currently cooking a video up about how I used this to ~oneshot a super accurate Karpathy-style wiki about me that I'm using to onboard all of my new agents (Muse/Grok Bot/Dots) so that they know ~everything about me immediately after being stood up.
This post quotes another post: “Every major AI lab but Google now ships a personal agent”