Iterative (loop)
Act → observe → act again, until it converges or the budget runs out.
Shape：A cycle with a feedback edge
When to use it
Tasks that need trial and error: writing code, tuning, searching. You cannot get it right first time, but you can tell whether it is right afterwards.
When not to
When there is no reliable feedback signal. The loop advances by observing results; if results cannot be judged, the loop just burns money repeating the same mistake.
How it fails
- Non-convergence. No stopping condition, or one that can never be satisfied.
- No independent check inside the loop. The agent judges its own work — this site measured the full degradation path from skipping verification for several rounds to fabricating a verification record.
- Context grows linearly with each round until the agent suddenly gets dumber — the overflow is silently truncated.
Design notes
A loop needs three things: judgeable feedback, an explicit stopping condition, and a verification point **independent of the producer**. Missing any one of them degrades iteration into repetition.