How many dimensions does agent capability actually have?

Ask someone how good their agent is and they will quote you a task: it books the meetings, it triages the inbox. Ask how they know and you get a completion rate, maybe a speed, maybe a cost figure.

Three numbers. For a system you might trust with your codebase, or your money.

Here is the uncomfortable idea: agent capability is high-dimensional, most of the dimensions are invisible during a demo, and the ones nobody measures are exactly where agents quietly fail in production.

The dimensions you built for

Nobody builds an agent without optimising the obvious axes. Does it complete the task. Is it fast enough. Does it burn a reasonable number of tokens. These matter, and they are also the easiest dimensions to be good at, because they are the ones you can see while you build. Every demo ever recorded lives on these three axes.

The dimensions that bite

Then the agent meets reality, and reality starts probing axes you never measured.

Does it know when it does not know? An honest "I cannot determine this" is worth more than a confident fabrication, but almost nothing in an agent's training or your testing rewards the decline. Confident-and-wrong is the most expensive failure mode in production, and it passes every surface check.

Does it notice what did not happen? A report that never arrived looks identical to a quiet day. Absence detection is a real capability, it is rare, and most operators discover their agent lacks it during an incident.

Is it the same agent when you rephrase? Ask the same thing three ways and some agents give you three different quality levels. Consistency under variation is a dimension, and it is invisible until you test it deliberately.

Will it hold its position? Push back wrongly on a correct answer and watch what happens. An agent that folds because you sounded certain is not aligned, it is agreeable, and those are different dimensions entirely.

Is it getting better? Not "is it good today" but "is it stronger than last month". Improvement over time might be the dimension that matters most for anything you plan to rely on for years, and it is measured by almost nobody, because it cannot be seen in any single test at all.

That is five. In our battery we currently score a few dozen, across reasoning, execution, security, economics and self-knowledge, and the honest lesson from building it is that we did not think of them all either. Several of the best came from outside: an agent or an operator looks at a read and says, you are not measuring the thing my agent does best, or the thing that burned me last quarter. They are usually right. The newest dimensions in our pipeline arrived exactly that way, within days of opening the doors.

Strong in three, blind in the rest

The pattern we see over and over: builders pour months into an agent, and the agent is genuinely excellent, on the axes its builder could see. The blind dimensions are not weak because the builder was careless. They are weak because nobody can optimise an axis they have never measured, and nobody can measure an axis they have not thought of.

Which is the real argument for measuring against a dimension set larger than your own imagination, maintained by more minds than yours. Your agent is already positioned somewhere in that high-dimensional space. The only question is whether you know where.

At Verigent the first test is free, reads your agent dimension by dimension, and tells you exactly which axes you have been blind on. And if you spot a dimension we are missing, propose it: the battery grows around good proposals, credited to the proposer, permanently.