A human team carries history no model fully sees. An AI agent team is, at least on paper, fully specified: models, prompts, tools, policies, topology. That makes it a natural test bed: if prediction cannot work there, it is unlikely to work for people.
The obstacles
Configuration is not behaviour. The same agent can behave differently on similar tasks, and model updates can silently change everything. Track R01 asks the prior question: are agent behaviours stable enough to measure at all?
A first experiment, planned
Run fixed configurations repeatedly on a task suite, record traces, and test whether simple features predict completion and cost on held-out tasks better than a naive baseline. If they do not, that is a result worth publishing too.
This is a concept note. It reports no experiments and no results.