simian-agent ships directed-mode chaos (plain-text intent → LLM-translated FaultManifest), an autonomous planning loop (health gate → topology snapshot → LLM-generated AttackPlan → bounded execution), four chaos engines (chaos-mesh, network-policy, envoy-fault, kube-state), and a safety stage at the executor chokepoint that enforces eligibility, blast-radius tier, duration, and concurrency caps on every fault before it lands.
Safe by default
Every fault flows through one chokepoint — the Fault Executor — which checks namespace eligibility, blast-radius tier, duration ceiling, and concurrency budget before any chaos driver runs. Cap violations surface as executor.rejected with a structured reason, not silent acceptance.
Autonomous or directed
simian chaos --intent "..." for plain-text directed faults. simian serve --autonomous for the LLM-driven planning loop with per-cycle budget caps, baseline health-gating, and clean LLM-down skip behavior.
Works on GKE Dataplane V2
Network chaos that the eBPF dataplane quietly drops fails its efficacy gate instead of being reported as applied. Where Chaos Mesh’s NetworkChaos can’t reach, the network-policy and envoy-fault engines work above the dataplane.
Install
Getting started goes from an empty project to Simian breaking Online Boutique, checking the damage and planning its own faults — on GKE in about 30 minutes, or on kind on your laptop in about 15. Then Deploying with Helm for your own workloads, or the Design doc for the architecture.