Iteration budgets

Every iteration is a model call, and every call carries the entire conversation so far. This means cost does not grow linearly with iterations — it grows roughly with the square of them, because each turn re-sends everything before it plus the new observation.

That is the real argument for a budget. A loop with no ceiling that gets stuck does not fail quickly and cheaply; it fails slowly and expensively, and the last iterations are the most expensive of all.

• Set the ceiling from observed need, not intuition. Run your tasks, see how many iterations the successful ones use, then allow some headroom.
• Count tokens as well as iterations. A loop with long observations can blow a budget in three turns.
• Treat hitting the ceiling as a signal worth logging, not just a stop condition. A rising rate of budget exhaustion means something changed.
• Return partial work when you can. An agent that stopped halfway usually has something worth showing.

The lab reports tokens and elapsed time for every run. Run the same task at different budgets and watch the token count grow faster than the iteration count — that gap is the cost of carrying the conversation forward.