iLogic is still alive: when a 50-line rule beats an agent

I build AI agents for Autodesk products. It is my job, and it is the part of my job that draws the most questions. Which is why it surprises people so much when my recommendation, after listening to a problem, is a 50-line iLogic rule. No LLM, no server, no tokens. A plain old rule.

It is not a contradiction. It is the consequence of an idea I repeat often: AI is the means, not the promise. And when the right means is simpler, the right means wins.

When the rule wins

iLogic wins when the logic is deterministic. If you can write the rule in one sentence without an "it depends" showing up, that automation does not need intelligence: it needs discipline. If the thickness is over ten millimeters, use template B. If the material is stainless steel, enable finish C. Known inputs, known outputs, zero ambiguity.

On that ground, the rule has advantages no agent can match. It runs inside Inventor, with no external services: it does not depend on a connection, a subscription or a vendor. It does exactly the same thing today, tomorrow and six months from now, and in engineering that has a name: reliability. It can be audited line by line, it ships versioned with the template that uses it, and any colleague with some VB can maintain it when you are not around.

When the agent earns its place

The agent earns its place when the exception shows up. Ambiguous inputs arriving in natural language. Decisions that change with context. Judgment that today lives in one person's head and gets applied one click at a time. There the rule falls short, because the rule only knows how to do what was already written.

The sentence I use to decide comes from my own Autodesk University session: a script automates the rule, an agent handles the exception. If your process is one hundred percent rule, an agent only adds cost and uncertainty: you pay tokens for something a conditional solves for free, and you accept a probability of error where there used to be certainty. If your process is full of exceptions that a person resolves by hand every week, that is agent territory.

Not a war, a division of labor

The mature conclusion is not picking a side. In the systems I build, both things coexist: the agent understands the intent, decides what needs to happen and delegates to deterministic logic everything that can be deterministic. The rule executes; the agent orchestrates and handles what the rule did not foresee. Each layer does what it does best.

What does not age well is the reflex of putting AI into everything. Every technology has its ground, and recognizing each one's ground is exactly the kind of judgment that separates a solution that works from a demo that impresses.

If you have a process on your hands and you are not sure whether it goes rule or agent, write to me. That fifteen-minute conversation usually saves months of the wrong project.