Uncertainty is an operating condition of agent systems: the language model’s behavior is probabilistic, its reasoning is opaque, and its failures can include hallucination, drift, and unbounded loops. Agent architecture must allocate resources within bounds even when the decision-maker itself is unpredictable. section 1.0 section 1.5
Three forms of uncertainty shape agent architecture:
- Output uncertainty: the same prompt can produce different responses. A retry may return a different answer rather than repeat the original result. section 1.2.2
- Behavioral uncertainty: the agent chooses actions at runtime. A reasoning-and-action loop can vary in both length and content. section 1.2.2
- Environmental uncertainty: the world can change between observations. Page layouts, codebases, and data sources may change or become unavailable. section 1.2.2
Conventional controls such as timeouts, retries, and tolerant parsers contain some uncertainty, but agent architectures must account for it from the start. Earlier distributed-systems patterns offer useful precedents: circuit breakers inform loop detectors, while saga transactions inform plan-and-execute workflows. section 1.2.2 section 1.5
















