for now, agents need rules

the same thursday taping had a twist i didn't see coming.

my guest built api security on behavioral machine learning a decade before this ai wave. back then his whole argument was that you can't secure apis by writing rules and signatures for them. the attacks change too fast. you need models that learn what normal looks like.

so i asked him the obvious follow up. does that argument apply to ai agents now?

i expected a yes. more ai to watch the ai.

he said no. the opposite.

ai agent control requires deterministic approach, which means rule-based.

his reasoning is clean. machine learning is probabilistic. it gives you a likely answer, not a certain one. an agent built on a probabilistic model will mostly do what you want. mostly. and if your control layer is also probabilistic, you've stacked two maybes on top of each other.

when what's at stake is an agent touching customer data, moving money or sending messages in your name, mostly isn't good enough. you need walls that don't negotiate.

he was careful about the time frame. he said maybe four, five, six years from now it looks different. but right now, rules are the answer.

that's a founder who bet his career on machine learning telling you not to trust machine learning to police itself. i'd listen.

he also told me where his new fund is looking, and it lines up. not at the models. at the boring barriers that stop enterprises from using them. security policies. compliance. cost control. data sanitization. designing the process around the agent so it can't wander.

and he said what a lot of operators feel and won't say on the record.

it's still the wild west when it comes down to ai security out there.

the numbers back up the worry. gartner predicted last year that over forty percent of agentic ai projects will be canceled by the end of 2027, because of rising costs, unclear business value or weak risk controls. that's not a model problem. that's a rules problem.

i run my own business on agents, so this hit close.

my agents research, draft, score and remind. they don't send. a human approves every outbound message before it goes. that's a rule, not a model. it doesn't depend on the agent being smart enough to know when it's wrong. it depends on a wall.

and the walls are dull. this agent can read the inbox but can't send. this one can draft a contract but can't sign. this one can see the pipeline but not the bank. none of that is clever. all of it is deterministic. you could explain it to a twelve year old.

that's the point. when something goes wrong, you want to know exactly why, and you want the answer to be a rule you wrote, not a probability you hoped for.

if you're a founder building with agents, the instinct is to spend all your energy on what the agent can do. spend at least as much on what it can't. write the list. put it in code, not in a prompt. a prompt is a suggestion. a permission is a wall.

and if you're an investor, the companies he described are worth a look. the gold rush is in the models. the durable businesses might be in the fences.

long term, i think he's right that this changes. the probabilistic tools will get good enough to guard each other. but the founders who survive the next few years will be the ones who didn't wait for that day to put walls up.

more on how i put agents to work without handing them the keys is in the ai book.

so pick the agent you trust most in your business... what's the one thing it should never be allowed to do, and is that written in code?