A systematic analysis of ReAct-style agents reveals they're burning 90.8% of their retry budget on errors that can never succeed — specifically, hallucinated tool calls to functions that don't exist. The study tracked 200 tasks and found the root cause isn't model accuracy but a fundamental architectural flaw: letting language models choose tool names at runtime through simple dictionary lookups like `TOOLS.get(tool_name)`. When models hallucinate non-existent tool names, systems waste precious retry attempts on guaranteed failures instead of recoverable errors.
This exposes a deeper infrastructure problem plaguing production AI systems. As agents become more complex — chaining multiple components for tool use, memory retrieval, and external integrations — reliability compounds in ways most teams aren't measuring. Industry analysis shows that even highly reliable components (99% each) quickly degrade system performance when chained together, dropping to 90% reliability with just 10 components. Most monitoring dashboards show acceptable success rates and latency while completely missing the efficiency massacre happening underneath.
The proposed fixes are structural, not prompt-based: classify errors before retrying, implement per-tool circuit breakers, and move tool routing into deterministic code rather than model outputs. This approach eliminates wasted retries entirely and reduces execution variance by 3x. The broader lesson extends beyond ReAct agents — as the industry builds increasingly sophisticated agent stacks, the gap between theoretical model capabilities and production system reliability will only widen without fundamental changes to how we architect AI infrastructure.