Eric Siu argues against relying on a single super-agent, instead building AI workflows with multiple specialized agents and strict guardrails. He uses a pick-N method for content: the system generates dozens of drafts, humans select which ones to publish, then tracks performance over 7, 14, 30, and 60 days. Every workflow requires human verification before anything goes live.
Chapters
One mistake that I made with my Open Claw agent initially, or my Hermes, I tried to have one super agent, right? It's also easier for your mind. It's easier for the AI as well, just when it has one job to focus on, right? What it's saying here is basically splitting up is kind of how we do things, right? We have a safety bar. Basically, like, you do want to set some type of criteria, and then you want to have some evals in place too, right? Eval is basically what the definition of success looks like, and you do want to have a safety bar too. And so our whole thing is like no reads, no writes, sensitive information should be covered up. You don't want to just let it YOLO and do whatever it wants, right? This one I think is unique too. We have this pick N piece, right? What that means is like for Short-form content, for example, if I want to scale that out, it'll give me 4 candidates. I can pick a couple and then it'll lock and load them and they'll schedule them. It'll measure the success of these 7 days, 14 days, 30 days, 60 days out. Same thing for SEO too. So for that autonomous engine optimization piece I was just talking about, you know, we'll have 40 candidates per week. I'll pick 20 and it'll publish 20 per week. ClickFlow, which is the content generator that we have, which has an API and an MCP, it'll write those WordPress, it'll push it to draft. I'll never edit a live page. Same thing for warm outreach when it comes to looking through our CRM. So 20 send ready per day. Shortfall is fine. Do not lower the bar. LinkedIn, you know, 10 per day and only if with verified funding, 12 months or higher. So this is the outbound piece, right? All these things like LinkedIn, for example, you can't easily just connect to an API. And so WordPress, you do need like an MCP for it. What we're looking for each day is like, okay, you still need a human in the loop here, you basically need an agent verifier here at the end of the day.