Matt won’t let anyone else on his team command his Slack bot. People ask him questions, and he types a message telling his bot to go get the answer for them. It’s a clunky arrangement, and he knows it.
His reason is sound. That bot is hooked up to so many systems that if anything went wrong, it could grab something inappropriate and hand it to the wrong person. Maybe somebody’s salary, sent to the wrong department.
You and I know not to share another employee’s salary, hourly rate, or hours worked. To an agent, Matt pointed out, that may not be obvious. Once agents start talking to other agents, which is where I think this is all heading, the problem gets bigger. My bot is named Kermit. Matt has AI Matt and AI Graham. When Kermit starts collaborating with AI Graham, who decides what crosses over?
The current answer is deterministic rules, plus governor agents that babysit the other agents and make sure they don’t break the rule set. Matt’s problem with that is simple. It’s hard to program in everything. The rules depend on who in the org has the bot, how senior they are, what work they’re doing, and they get complicated fast.
The bitter lesson, in a nutshell
If you don’t know the bitter lesson in AI, it’s worth some research. My short version: raw process and compute is going to overwhelm specificity almost every time.
A couple of years ago Google put out healthcare-specific models. By the time they were done, the frontier models, with all their additional neural connections and nodes, simply outperformed them. More information is just better for LLMs.
Matt said he’s found it fascinating that we haven’t seen more models come out differentiated on a single category, like law or x-rays for cancer. The acceleration is coming from generalized intelligence.
I’ve watched this happen to my own work
Longtime listeners know I had N8N flows going all over the place. Deterministic workflows, connecting this to that, soldering things together.
Then in January it was like a switch flipped. Matt says OpenClaw, and yes, but what really did it was Claude Code becoming so much more useful. All those complicated things we were doing got wiped away by the ability to create edge functions or Cloudflare workers that did the same jobs. Process overwhelmed the former approach.
It’s happened again with orchestration agents like Dot, Muse, and Duet. They bring so much process to bear that they solve problems my old workflows needed hand-wiring for.
Why your governance rules are next
At this point in time you need a stack of deterministic rules so you don’t end up in trouble. I’m not arguing otherwise. That stack is a point in time.
What comes next is something like angel-on-the-shoulder agents. Their prime directive is interpreting what Matt’s bot is supposed to release, and they act as a proxy on his behalf instead of relying on a pile of rules. They’ll be better at it, because they’ll simply get that Matt is the founder and the CEO, and that some things obviously stay private. Then they’ll start regulating the other agents.
Matt pushed back, and his objection matters. He’s a pretty unique guy. He’ll share a call recording that most CEOs wouldn’t, and he gives his bots a lot of credit cards. Plenty of people hold their cards closer to the chest. Can a bot learn the difference between those two personalities, call by call? Was something private said? Is that bad or not bad?
My answer is that you don’t need to automate every decision away. My clients at bigger companies already have executive assistants and chiefs of staff having those conversations all day: this priority, that focus, what gets communicated. They’re just happening with a person. In a smaller company, you’re your own voice, so the bot pushes back. Hey Matt, this conversation looked like it included private details about Kevin’s life. Want that included? No, strip it out. Done.
Matt is going for full automation anyway. He says he can see how much of his day is answering things a bot could handle if it knew it had permission. He just doesn’t trust the guardrails yet. His team is also tech-savvy enough to break the rules. One of his developers sent him a screenshot of his bot conveying a bunch of private information in a DM. Matt thought his bot was locked down. The developer had generated the screenshot.
My own view: the rules are a bridge. Getting from rules to context-aware governors is going to be fascinating, and it’s going to be ragged to get there.
