Cascade resistance

Cascade resistance scores, on a 0-to-1 scale, how well an architecture keeps one fault from escalating into a whole-system failure. It is a property of the wiring, not of any single agent: hold the agents constant, change the topology, and the score moves. What makes it useful is the ranking it forces across common patterns:

• A chain resists least, since any uncaught step-level error rides straight through to the output.
• A supervisor that validates each result before accepting it buys resistance by inserting a checkpoint between hops.
• Fan-out with an orchestrator-worker pool localizes a bad worker when the orchestrator refuses to merge unverified state, and leaks it when the orchestrator does not.
• Independent consensus voting withstands a lone fault best, because an uncorrelated minority error gets outvoted, at the cost of extra model calls and latency per ballot. Debate is conditional: independently seeded dissent can argue a lone fault down, while a correlated persuasion loop amplifies it into a shared error (see multi-agent debate failure).

The score turns blast radius and containment rate into one number you can defend in a design review, while the failure-mode pillar maps each pattern to the faults it survives. Chains fail fastest; consensus resists longest, at real cost in tokens and latency.