Agentic tasks amplify model quality differences, because an error at step two poisons every observation after it. A stronger model that needs six steps is frequently cheaper end to end than a weaker one that needs twenty and still fails.

University of Michigan - Ann Arbor

A topic of the Building AI Agents course. Planned for 1.5 hours at Easy difficulty.

Choosing a Model for an Agent

Capability Versus Cost

For agents, the deciding capability is rarely raw knowledge; it is whether the model reliably respects tool schemas, stop conditions and formatting contracts across dozens of turns. A model that drifts from its contract at turn fifteen cannot be deployed unsupervised.

Learn Before

Related