How do you choose the right AI model for an agent?
Choose the language model per task and per role: a fast sprinter for simple work, a workhorse for everyday work and an Einstein for heavy thinking. Evals show which model is good enough on accuracy, cost and speed.
Explanation
One agent often uses all three: the sprinter handles intake, the workhorse the regular work, and only the hard cases go to the Einstein. That keeps costs low without losing quality.
When the inference bill rises, model choice becomes a trade-off between accuracy, cost and speed. Evals help you decide when to optimise, switch to another model or train your own.
An eval is a test with real cases for which you know the right answer: the golden set. Run it with every change and you see straight away whether a cheaper or newer model can do the job.
Provenance counts too. Open-weight models from the US, China and Europe can be downloaded freely and run yourself. That gives control over data and costs, and at the same time turns a model's origin into a security question: who trained it, under which licence, and where does it run? Those details belong in the agent passport.
What it looks like in practice
A customer service agent sorts messages with a sprinter, answers standard questions with a workhorse and sends complaints with legal weight to an Einstein, after which an employee reviews them.
