Fault injection is a testing technique in which engineers deliberately introduce a problem and observe what the system does next. In a multi-agent workflow, they might replace one agent's response with incorrect data, interrupt a tool call or simulate a missing document.
The objective is to see whether the workflow notices the fault, corrects it, asks for help or carries the error forward. Testing controlled failures can reveal weaknesses that normal success-path demonstrations miss. It also helps evaluate timeouts, retries and fallback strategies.
Faults must be designed carefully. Randomly breaking components may create dramatic failures without teaching much about realistic risk. Tests should match plausible operating conditions and distinguish a model reasoning error from an infrastructure failure or deliberately hostile input.
Run fault injection in a safe, authorized test environment with monitoring and rollback controls. Record the injected condition, observed behavior and recovery outcome so the same test can be repeated after a workflow change.