Fault Tolerance and Liability Apportionment - Akash Narayan

ABSTRACT
When a chain of agents and APIs produces a loss, two bodies of expertise arrive and talk past each other. Reliability
engineering treats the event as a fault to be contained, and reaches for retries, circuit breakers and compensating
transactions. Law treats it as a question of attribution, and reaches for defect, negligence and the allocation of provider
status along a value chain. This paper argues that both under-perform on agent chains for one shared reason: their central
assumptions were built for components that fail visibly and stop. An agent that has gone wrong does not stop. It returns
output that is well formed, confident, plausible and incorrect, which is closer to a Byzantine fault than a crash, and neither
the engineering toolkit nor the doctrine of defect has an affordable answer to Byzantine faults. From this the paper draws
a claim about where apportionment will actually happen. The cost of reconstructing causal share rises with the number of
handoff pairs while the value of a typical individual loss does not, so beyond a shallow depth after-the-event
apportionment costs more than the loss it resolves. Liability will therefore be allocated in advance, by protocol and
contract, in the manner the card networks used for counterfeit fraud, and allocation rules will attach to
machine-observable facts rather than to litigated findings. The paper sets out what a chain must emit for such rules to
bind, argues that bounded and attributable loss is a more useful design objective than low error rates, and treats the shared
model substrate beneath nominally independent parties as the systemic risk that the insurance market has already begun to
price by withdrawal.

Keywords: multi-agent systems; liability apportionment; fault tolerance; Byzantine faults; agentic AI; product liability; systemic risk;
blast radius; attribution; API value chains

Author: Akash Narayan