Propagation radius (agent systems)

Propagation radius is the count or fraction of downstream agents and steps one fault reaches before it is contained, read across a specific topology rather than reported as a single global figure. The operational move that makes it concrete: fix the injection point, then count reached nodes as a share of everything reachable from it. In an orchestrator with N workers, a poisoned worker output the orchestrator trusts can reach every sibling that reads shared state; in a chain of depth D, an uncaught error at step k reaches the remaining steps. Same fault, different radius, because the graph differs.

Radius is measured per fault and per topology, so it is reported as a distribution over injection points, never one headline number. It sets the numerator that containment rate works against and the exposure that cascade resistance scores at the architecture level. Radius reads a fault’s breadth, how much of the topology it touches, while its depth counterpart, propagation depth, counts the successive hops the same fault survives before a check contains it.

Engineers coming from site reliability practice sometimes call this a system’s “blast radius”, an informal borrowing for the same idea. Propagation radius is the version you can measure: a per-fault, per-topology count of the reach of error propagation, the mechanism that carries one agent’s mistake into the agents acting downstream of it. Where SRE usage stays informal, this lane gives the quantity a name it can measure, and that lets a review compare two architectures on one axis instead of trading impressions about which design feels safer.

How wide the radius runs depends on the boundaries a fault is allowed to cross. A checkpoint that validates a worker’s result before the orchestrator merges it, or an MCP gateway that mediates the tool and data calls passing between agents, caps the reachable set, so nodes on the far side of an enforced boundary fall out of the count. That is the lever a design review has: move a boundary inward and the same fault reaches fewer nodes. Shrinking that reachable set is the entire job of containment. See the multi-agent failure-mode taxonomy for where propagation radius sits among propagation metrics.