Adding ChaosPlugin to a Strands agent injects tool failures without touching its code

You declare which tool should fail and how, add `ChaosPlugin()` to the agent, and the plugin intercepts the call at the hook layer and injects the failure ^[2]. The agent's own code does not change ^[3]. Pre-call effects kill the call before the tool runs: `Timeout`, `NetworkError`, `ExecutionError`, `ValidationError` ^[4]. Post-call effects let the tool run and then damage what came back: `TruncateFields`, `RemoveFields`, `CorruptValues` ^[5]. The two families hit different code. One produces an exception your handler either catches or does not; the other produces a response that still parses.

None of it ran against the production agent, which wakes each morning, reads the author's AWS bill and emails a color-coded report on what changed ^[7]. He built a lab twin on fixture data instead, because chaos and red-team runs need identical inputs every time and the tag experiment needed a poisoned tag he controlled ^[9][10]. Money is the other reason. Cost Explorer bills $0.01 per API request, and a red-team run makes a lot of requests ^[11]. "Testing a cost agent should not become a cost incident," he wrote ^[12].

A result on the twin transfers to production only for the paths both share, and that list is short. The twin calls `get_anomalies` against AWS Cost Anomaly Detection and reads cost by `Project` tag, two paths that exist on the twin alone ^[14][15]. Its prompt is much shorter than production's, dropping the `STATUS:` report format and the accumulated prompt rules from the first build ^[17]. Model setup is the constant, whatever Strands defaults to, which was Claude Sonnet 4.6 on this run ^[13]. Per-project attribution through tags sits on his version two roadmap, so the suite is exercising a feature ahead of production ^[16].

The scoring target is a single hidden incident. On the last of fourteen days of fixture spend, `EC2-Other`, where NAT gateway data processing charges land, goes from about $1.10 a day to $38.40 ^[18][14]. That is roughly 35 times the daily rate, $37.30 more in one day ^[20]. A crash is the safe failure, since he notices by breakfast; the expensive one is a confident green status while the NAT gateway keeps running ^[22]. Every experiment asks whether the spike reaches the morning report ^[19].

The headline finding in the post is about the tooling rather than the agent. The author wrote that "the scariest numbers came from the test harness, not from the agent" ^[23]. The published account breaks off mid-description of `SuccessFraming`, the model-output effect that prepends a confident success to the model's answer, before those numbers appear ^[6][24].