| Architecture Pattern | Error Fallback Route | Max Retry Limit | Token Explosion Risk | Production Resilience |
|---|---|---|---|---|
| Unbounded Loop (Implicit State) | ❌ Missing (Loops to self) | Infinite (Unbounded) | HIGH (Token Exhaustion) | Fragile (0%) |
| Directed Graph (Explicit Map) | ✅ Explicit Branch Edge | Static Conditional | MEDIUM (Bounded Loop) | Resilient (65%) |
| Guarded Harness (Policy Rails) | ✅ Guarded Fallback Node | Dynamic Guard Constraint | LOW (Capped Budget) | Guarded (100%) |
When a tool call or evaluation fails, direct the execution token to a safe recovery path instead of re-entering the naive planning loop.
In a naive unbounded loop, an API error cycles directly back to the LLM planner with no recovery boundary, consuming tokens indefinitely. Selecting Guarded Harness inserts policy rails. The execution token reaches Post-Guard, which enforces validation and a retry threshold. Each repeated failure charges the retry counter. Once it hits the limit of three, the circuit breaker triggers safe degradation and locks recovery at one hundred percent.