Six failure patterns — each a distinct mechanism by which a local failure becomes a system-wide event.
Pattern 1
Silent Error Propagation
A failure produces structurally valid but semantically wrong output. Downstream agents process it as if it were correct — no error signal is emitted at any hop.
Pattern 2
Error Amplification
Each pipeline stage compounds the error from the previous stage. A small initial distortion becomes a large final distortion, because agents process and re-interpret bad input rather than simply passing it through.
Pattern 3
Runaway Feedback Loops
An agent's output directly or indirectly feeds back into its own inputs, creating a self-reinforcing cycle. Without a circuit breaker, the loop runs until external intervention or resource exhaustion.
Pattern 4
Resource Exhaustion Cascade
One overloaded or slow agent causes a queue of pending requests to build in dependent agents. Retries multiply the load. The overload spreads upstream and eventually brings down agents that were initially healthy.
Pattern 5
Dependency Chain Collapse
A critical shared dependency fails, and all agents that depend on it fail in sequence. The more agents share the dependency, the larger the simultaneous failure footprint.
Pattern 6
Adversarial Cascade
An attacker deliberately causes one agent to produce malicious output, using that agent as a stepping stone to reach other agents they couldn't attack directly.