A cheap model call can become expensive when someone has to rebuild the project around it. Long-running work pays for repeated context, failed attempts and recovery. A model upgrade costs something too if the new worker needs weeks of explanation before it can continue.
Keeping a project coherent costs compute, and that cost competes with new work. In our September study, durable agents used 82% fewer tokens over 42 days and needed 13.7 hours of human time per programme, against 41.6.
Measure the undertaking
We count the whole life of a project: productive work alongside waiting, repairs and model changes, and the time people spend stepping in. A recovery counts only when the lost ability is back and stays useful.
Different abilities need separate measures. A team can stay busy while losing verified knowledge or getting worse at coordinating, so we track those alongside the continuity of its plans and the evidence behind its decisions.
The economic promise is that experience lowers future effort. A tested method saves work across many projects and successor models, so durable systems spend more of their resources on hard new problems and people spend less time keeping an AI project on course.