Runaway Loop Prevention

Agents fail by looping more often than by acting wrongly. A step produces output that fails a check, the agent retries, the retry produces the same output, and the cycle continues consuming tokens and time until something external intervenes. Detection is easier than it appears — repeated near-identical requests, a step count exceeding expectation, elapsed time past a threshold — but only if something is watching, which is why loop prevention is a platform responsibility rather than a prompt instruction. The model cannot reliably detect its own repetition.

SkillTokenBudget bounds what a single execution may consume before it runs, which converts an unbounded loop into a terminated one. The Interaction Log records each execution separately, so a repeating pattern is visible as a sequence of near-identical entries rather than as an unexplained cost spike at the end of a month.