When Agents Fail

Agents get stuck: the same failing tool call, retried forever. No exit, no progress — and every spin costs a full model call against a growing context. Teams have watched agents retry a dead endpoint dozens of times until the token cap saved them — burning dollars to produce nothing.

Every agent needs hard budgets:

• Max steps — 10–25 covers most tasks
• Token/cost cap per run
• Wall-clock timeout
• Loop detection — same tool + same args twice = kill