Build a proactive agent

I've a researcher agent. I call him Scout. His job is to research companies.

Each day Scout takes a handful of companies and researches them. He builds a picture of each one: who actually works in the marketing department or team and what their roles are, what their leadership is saying in public, what they run in market, their ads, the search terms they show up for and the ones they buy, what their customers say in reviews, and what their competitors are doing beside them. Then he writes it all into the CRM, so the research lives where the work happens, not in a folder nobody opens.

Making an agent do this well is pretty similar to people management, and you need the same patience you'd give a capable new hire. I spent a full day clarifying the details for Scout, his manager (not kidding) and for myself. What exactly is his goal? Where does his output go? What does he depend on? What tools does he have and what is he allowed to touch? Who does he escalate to? What happens when he's blocked? The robots will happily skim stuff that sounds good without thinking it through, if you let them. When I pushed at one stage and asked where the actual data was, I found I'd been given a summary, the way a person will give you a summary if you never ask to see the working. The fix: every claim carries its source, and the reviewer (me) can open the evidence.

Agents work brilliantly with number goals, but only when the number comes with a quality goal attached. Scout's goal is a fixed count of excellent pieces of research a day, and excellent is defined: every source checked or honestly marked as not checked, every person verified as actually still in their job, every claim linked. There's a second agent whose only job is to sample the work daily and follow the links. When the checks come back clean, Scout's capacity scales up automatically. When a defect appears, it scales back down, and the lesson gets written into the method so the same mistake can't repeat. I never have to revisit the decision. That was the design rule I cared most about: I'll happily spend a day setting a role up properly, but I will not come back every fortnight to manage it.

And that's the proactive bit. A well-designed agent never stops to ask permission for work. Scout researches as far as he can go, flags what he can't crack, routes what he finds to the people who need it, and only ever stops at the few hard gates every business should have: spending money, sending things to real people. If a rule blocks him, he doesn't sit quietly being blocked. He proposes the change, every day, until someone decides. I stole that from managing humans. It kills me to see capable people waiting politely.

If you're thinking about doing this yourself, the honest advice is: the writing of the role is the work. What's the goal, in a number and a standard. What are the outputs, countable. What does he need, named. What can he touch, exactly. Who checks him, and what happens when the check fails. That's a job spec, and it's the same discipline whether the new starter is a person or a process. The difference is that once you've done it properly for an agent, it holds.