A production agentic system sometimes enters infinite loops, calling the same tool repeatedly with identical parameters. QA testing confirms this happens in roughly 3% of runs. What architectural safeguard should be implemented?

Infinite loop detection and prevention must be implemented in the orchestration layer (code), not trusted to Claude's instruction-following alone. The safeguard has two components: active loop detection — compare each tool call against recent call history and detect identical tool-and-parameter pairs; and maximum iteration count — enforce a hard cap, such as 50 steps, with human escalation or graceful failure when reached. When a loop is detected, inject a guidance message such as: 'You have called this tool with these parameters 3 times without progress. Consider an alternative approach.' The larger-context approach is irrelevant because Claude can already see its call history in context and still loops; the issue is a reasoning failure, not context blindness. The temperature-zero approach makes loops more likely, not less, because a deterministic model will repeat the same mistake every time with zero variation. The instruction-only approach is insufficient because agentic loops occur when the model genuinely believes each next call will differ.