In every bottleneck, an opportunity.

You study how AI coding agents behave, and benchmarks only take you so far. A leaderboard tells you which model solved the most puzzles. It doesn’t tell you which one plans the better architecture, or which one burns a hundred thousand tokens second-guessing a three-line fix. You want to watch them do real work, one step at a time.

So you build a custom workflow made for the experiment in Expecto Centum—just the stages you want to study, each carrying its own agent and model. You write one ticket, run it through, then run it again with a single change: a different model on the coding stage. Everything else holds still. The Dashboard’s token counter breaks each run down stage by stage, so you see exactly where the extra spend went. The branch each run leaves behind shows you what that spend bought.

By Friday you have a dozen runs of the same ticket, a spreadsheet of per-stage token counts, and a finding with an actual control group behind it.