ARTICLE · CONTROLS FOR AI WORK · PART 4 OF 6 · Interlocks · guarding

Funnels, cones and the prime directive

Every rule you give an AI agent is one of two things. A funnel shows it where the work should go and lets it find its way there. A cone is a hard line it must never cross. Most people write all their rules as cones, in one flat list, with no order. That’s how you get an AI agent that’s boxed in on the small things and still not safe on the big ones. Get the difference right, rank them, and keep the hard rules few.

The same questions apply to a sales pipeline, a purchase order or a month-end close: what stays fixed, what needs judgment, and who decides?

What the AI field has found

Three findings point the same way.

Too rigid and too vague both fail. A widely cited guide to context engineering says instructions need to be at the right “altitude.” Too low, and people end up “hardcoding complex, brittle logic in their prompts.” Too high, and they give “vague, high-level guidance that fails to give the LLM concrete signals.” Too rigid is the parking lot full of cones. Too vague is the persona that fades.

More rules, less following. The IFScale study gave AI models more and more instructions at once. The more instructions, the fewer got followed.

Priority isn’t automatic. AI systems are supposed to rank the developer’s instructions above the user’s. Two 2025 studies found they don’t do it reliably; one found models were swayed more by a user sounding confident or authoritative than by which instruction was supposed to win. If priority matters, you have to state it, and for the things that truly can’t happen, you can’t rely on the instructions alone.

And on the hard rules themselves, a widely used practical guide to building agents says actions that are “sensitive, irreversible, or have high stakes should trigger human oversight until confidence in the agent’s reliability grows.” Its examples: cancelling orders, authorizing large refunds, making payments.

From the shop floor: guarding, and what gets to stop the robot

When you design guarding for a robot cell, you start with time and distance. How fast is the robot moving? How long does it take to stop? How close can someone’s hand get before a sensor sees it? The robot isn’t allowed to reach the guard, because hitting the guard is the same as hitting a person. So you work it backward: the sensor has to be far enough out that the robot can stop in time. There’s a standard way to calculate that distance. Nobody guesses.

That stop decision happens in milliseconds. There’s no time to weigh eighteen variables. It has to be a simple check that outranks everything else in the program.

Compare that with inspecting a part on the same line. If a part comes every thirty seconds, the vision system can take three or four seconds to weigh eighteen measurements and decide good or bad. Same machine, two very different kinds of decision, each sized to the time it has.

The other lesson from the floor: the safety circuit outranks the program, and the program outranks anyone’s convenience. Nobody holds a meeting about whether the light curtain matters more than cycle time.

With a robot, the thing I worried about was pinning someone against a conveyor. Almost anything else can be fixed. That one can’t.

A prime directive

Elon Musk described the same idea for Tesla’s Autopilot in a 2021 interview with Sandy Munro, as CleanTechnica reported it: “the prime directive for the Autopilot system is don’t crash.” He added, “That really overrides everything.”

What makes that work isn’t the rule itself. It’s that it’s first. The system isn’t balancing fifteen equal rules; it checks one thing that outranks the rest, and it can check it in an instant. Then everything below it gets a priority too.

The same idea shows up in business. Jeff Bezos split decisions into two kinds in his Amazon shareholder letters: two-way doors, which you can walk back through, and one-way doors, which you can’t. Move fast through two-way doors. Take care at one-way doors.

How it carries over: funnels, cones, and a ranked order

Most rules should be funnels. They point the AI agent in the right direction and leave it room. I once set up a robot to push a wire pin into a hinge. Clamped tight and driven to the same spot every time, the pin rubbed and bound. Once we let the head float a millimeter, the hole in the part guided the pin to center. “Propose a sensible default before asking a question” is a funnel. “Show the arithmetic” is a funnel.

Cones are for one-way doors only. Sending, paying, signing, ordering, deleting, publishing, anything that could hurt someone. Those are the interlocks. Keep them few, so each one stands out.

Rank them. Say what comes first, and what wins when two rules pull against each other. A flat list makes the AI agent weigh everything against everything on every decision.

Make the top rule quick to check. Like the robot’s stop, the most important rule should be something the AI agent can check instantly. A rule that needs a paragraph of judgment can’t be rule one.

Here’s how the A.N. Creation Office does it. Its Project Manager Bot has about two dozen principles, nearly all of them funnels. Three are marked above the rest:

Three rules are never relaxed for speed: Never invent facts, Trace every spec to its source, and Severity first.

And its one-way doors are written as their own principle:

A reversible decision (a two-way door) is made fast, by the inventor, with what is known. An irreversible one (a one-way door) gets care: signing a contract, ordering a mold, spending money that cannot come back, a first public showing before a patent decision, anything that can hurt someone.

The real locks are outside

One honest limit. A written rule makes the AI agent more likely to stop at a one-way door. It doesn’t make it certain. On a machine, you don’t trust the program alone for safety; you add guards and safety circuits that work even if the program fails.

For an AI agent, the guards are the permissions you give it. If it shouldn’t send email on its own, don’t give it the ability to send email on its own. Give it only the access the job needs, and make a person the last step on anything that can’t be undone.

Try this

Take the rules you’ve given an AI agent and sort them into two piles. Funnels: guidance on where the work should go. Cones: things that must never happen. If the cone pile has more than a handful, most of them are funnels written as cones; rewrite them. Then put the cones in order, and write one line saying which wins when two conflict.

Sources

These pages sit with the rest of the Advance Now family. The hotbotspot tool library writes the methods up for AI. A.N. Creation Office is where the public AI agents run as open experiments.