Fault injection (agents)

Fault injection (agents) is a controlled experiment: you introduce a chosen fault at a known point in an agent topology and measure how far it travels before something stops it. The known point is what makes the result attributable: you recorded where the fault entered, so its downstream effect traces to a cause you chose.

The technique is decades old in software testing; the agent adaptation moves it to a topology of coordinating agents, where the fault of interest is a poisoned message or a silently stale tool result.

A shallow fault-tolerance test breaks something, watches whether the system recovers, and logs a pass. The rigorous version reads two quantities instead: the reach of the fault and the fraction held to a single hop, each reported with an interval.

It is the primitive, not the practice. Chaos engineering for agents is the broader discipline of running such experiments to build confidence in a system; fault injection is the single act of seeding one fault so that discipline has something to measure.

Which faults to seed, where, and how to read the propagation is the work of rigorous agent-reliability testing.