AI Agents Still Need Too Much Babysitting
By now you’re likely very tired of hearing about how someone’s 10x’d their productivity by using AI agents, or how they’re automating 90% of their work through elaborate agentic graphs or looping.
I really love a lot of AI tools, but the AI hype bros get something wrong: AI is not nearly as “set and forget” as they claim. There’s this notion that they’re basically autonomous - with the right skills, guardrails, and prompts, you can just let them go and do their thing. Unfortunately, that’s not the reality. At least not yet.
Grok Bot came out a few days ago and it is one of the most impressive tools I’ve used. It’s one of the few “it just works” agentic tools that exits. However, it’s held back by lack of customization (you don’t get to pick which models you use) and a lot of it is held in proprietary clouds, so you have a case of true vendor lock-in, which is not my favorite. Otherwise, though - it’s multiplatform, each agent gets its own cloud virtual machine with a desktop that can run terminal commands, a browser, and install software. And the agents you use effortlessly communicate with each other.
Whether you use a harness like Claude Code, Hermes, OpenClaw, or Cursor, you can’t go wrong in most respects. But Grok Bot’s UX was so good. No tinkering, no arduous, tedious setup to get it really up and working.
Did I still have to babysit them more than I’d like? Sure. But I think it’s time to come clean and admit a human needs to be in the loop still - to start / initiate something meaningful, and then for polish and taste - the last remaining bits before an output can be considered complete.
If you don’t want to learn about it, I don’t blame you. I’d rather do it for you.
This bog post is a stub, and 0 words of it were generated with AI.
I want to update it later with something I’m working on. More to come.