Error propagation (multi-agent)

Error propagation is the process by which one agent’s error is passed downstream as trusted input to the agents that act after it, so a single local mistake compounds across the system rather than staying at its source. Two shapes matter. Along a sequential chain the error deepens: every stage builds on the corrupted state it inherits, and the output drifts further from correct with each hop. Across a fan-out it widens: one bad result is consumed in parallel by many peers, correlating their failures at once. Where that same shared cause sits behind the steps of a chain, an end-to-end rate is worth reading with the common cause priced in instead of assumed away.

Propagation depth, the number of successive hops an error survives before a check catches it, is the measure that turns this from anecdote into a quantity you can bound. It also fixes the term’s place among its neighbors: propagation is the mechanism, propagation radius is the reach that mechanism achieves on a given run, and cascade resistance is the wiring’s standing tendency to damp or amplify that reach. The multi-agent failure-mode taxonomy maps which failures propagate and which stay put.

Error propagation and cascading failures

Error propagation is the mechanism behind a cascading failure: the cascade is what the propagation looks like once it has run its course across the topology. A single well-formed but wrong output clears validation, is trusted by every agent downstream, and the corrupted state keeps spreading until a check or a boundary stops it. The cascading-failures reframe treats that spread as an engineering quantity rather than a security incident, splitting the question into how far each fault reached, its propagation radius, and, across faults, the fraction held to a single hop, the containment rate.

How error propagation is measured

Error propagation is measured by fault injection: a chosen fault is seeded at a known point in the topology, and its downstream effect is recorded rather than guessed. Two quantities come back, the depth an error reaches along a chain and the number of peers it corrupts on a fan-out, and the independent-steps reliability calculator folds the per-hop containment behind them into an end-to-end reliability estimate for independently failing steps. Where many agents reach their tools through a single MCP gateway, that gateway is both the chokepoint one fault can travel through and the natural boundary at which to check it. What a propagated failure spends against the service’s own tolerance for them is what the SLO and error-budget calculator prices. The tolerance it spends against is an error budget, the count of failed tasks a completion-rate target allows over a stated volume and window, which is why a wide propagation radius is expensive rather than merely untidy.