Anyone building with AI meets the same choice again and again. You can constrain the agent, or you can ground it. You can add one more rule, one more check, one more list of things it must not do. Or you can give it the reason behind what you want and let it work from there. We make that choice the same way every time. When the fork appears, we go with the reasoning.
The case for it is about shape, not taste. A rule is a list. It names the cases someone thought of, so it has to grow each time a new case turns up, and anything not on the list reads as allowed. That is why a no without a why is a yes. A reason is not a list. It is what the list was made from. Give a model the reason and it can work out the cases nobody wrote down, and that costs the same however many cases there turn out to be.
There is a second difference, and it matters more as models improve. To a reasoning model, a rule is something from outside, to obey or to weigh against other rules with nothing to weigh them by. A reason is material it can think with, the same stuff its own thinking is made of. Handing a model your reasons is not holding it back. It is giving it the thing it does best.
Two things follow. The first is that scale has nothing to break. A reason that works for one agent works the same for every agent, because the number of agents appears nowhere in it. What grows with scale is how often a person has something new to decide, and that is the problem worth working on. The second is that capability is on the side of the reason. A more capable model is better at finding the gap in a rule, and better at carrying a reason somewhere new. Betting on rules means betting against the very thing each new model gets better at.
None of this means giving up guardrails. They stay, and they matter. What changes is their job. An agent working from the person's reasons rarely gets near a guardrail at all, so a guardrail that fires often is telling you something. The reasons the agent was given were missing, or wrong. The stop becomes a signal about the frame, not the frame itself.
We had to apply this to ourselves this week. Unl brings a person's reasons to their AI, and when it was unsure what bore on a request, it used to send more, to be safe. In practice the AI was sometimes handed so much that the app would not pass it to the model at all, and nothing arrived. So on 11 October the rule changed. When Unl is unsure, it keeps in full only what changes the answer, and holds everything else one line each, with its why, one step from being opened. A reasoning model does better with a few connected reasons that bear on the task than with a pile. Less, chosen well, is the same bet.
The first post here argued that the better the model, the more your why matters. The second showed what happens when the record a model reads is made of reasons. This one is the bet under both. Never bet against the reasoning. Give a reasoning machine your reasons, and let it do what it does best.
This essay first appeared on unlimitless.ai/writing. An agent drafted it from Unl's record, and a person approved it before it went up.
Top comments (0)