Containment rate

Containment rate is the fraction of an injected fault a system holds at the boundary where it originates, before the error crosses into the next agent or step. It answers a narrower question than whether the run passed: the fraction of injected faults held to a single hop (the per-hop containment probability c), reported with a confidence interval.

Because a fault either clears the first hop or it does not, containment is a proportion, and a proportion needs an interval. The flagship profiler is designed to seed the same fault across many trials and report the contained share with a bootstrap confidence interval, so the estimate carries its own uncertainty. The profiler is unshipped: a containment rate for your own system is one you measure yourself, and the interval convention it will follow is the one the studies we have published already apply to every rate they report.

It mirrors propagation radius, which supplies the population a fault could have reached, and it aggregates into cascade resistance once scored across a full topology map. Measure it before trusting any single pass rate; a containment number without its interval is a claim rather than a result.

How to calculate containment rate

Inject a known fault at one point, run many trials, then divide the trials where the fault stayed at that hop by the total injected: contained faults over injected faults. Because the result is a proportion, report it with a confidence interval instead of a lone percentage, following the same reasoning as an eval confidence interval; the interval narrows as trials accumulate. The place you enforce and then measure first-hop containment is a chokepoint in the agent graph, such as a validator, a supervisor step, or an MCP gateway that screens tool calls before they reach downstream agents. To compose per-hop containment across a whole chain, the multi-step reliability calculator does the multiplication for you.

Containment rate vs resolution rate

Resolution rate measures a different endpoint: the share of runs, or in a support bot the share of conversations, carried all the way to a correct final outcome. Containment rate measures one boundary, the share of faults held to a single hop. A system can contain most faults yet resolve few tasks, when the surviving faults land on the hardest steps; it can also resolve many tasks while containing poorly, when downstream agents happen to correct an upstream error. The two answer separate questions, so a rigorous report tracks both: containment for how far a fault spreads, resolution for whether the work got done.

Containment rate vs deflection rate

Deflection rate counts contacts kept out of a human queue: the share of sessions a bot handles end to end without escalation. It is an operations metric about human workload. Containment rate is a propagation metric about where a fault stops. In a support setting the two often move together, since a contained failure tends to avoid a handoff, yet a deflected session that ends in a wrong answer is a containment failure the deflection number hides. Reporting deflection on its own rewards suppressing escalations, so pairing it with containment keeps the quality of those deflected sessions in view.

Chatbot containment rate benchmark

Customer-support automation uses the same phrase for a different quantity: there, containment rate is the established headline metric for the share of conversations that end without a human takeover, close to what this page calls deflection rate. Teams often want one benchmark number, yet containment swings widely with the intent mix, since password resets contain near the ceiling while billing disputes sit far lower, so a single blended figure across a whole product says little. A defensible benchmark segments by intent, reports each segment’s rate with its confidence interval, and holds the trial count steady across comparisons. The discipline this lane applies to injected-fault containment carries over directly: quote each segment’s rate with the interval that bounds it, so the benchmark reads as a measurement.

Containment rate and propagation radius are the paired quantities that turn a cascading failure into something you can measure per topology. The fuller account of how one well-formed but wrong output becomes a cascade, and which levers raise containment at each hop, lives in cascading failures in agent systems.