ChaosAgent: Multi-Agent Fault Injector
Autonomous Chaos Engineering & Byzantine Fault Injection for LLM Agents
Role
Lead Systems & Security Engineer: fault injector harness, Byzantine interceptor, resilience benchmark runner
Primary Stack
Python · LangGraph · FastAPI · Docker · Pytest · Prometheus
“Production multi-agent systems suffer from cascading failures when tool APIs hang, prompt drift introduces hallucinations, or Byzantine agent nodes return corrupted outputs.”
Must inject 6+ discrete failure modes: latency spikes, dropped tokens, hallucinated tool schemas, partial network partitions
Automated resilience scoring and recovery time objective (RTO) benchmarking
LangGraph cyclic agent loop compatibility with zero production downtime
Why deterministic fault injection at the SDK interceptor layer?▼
Injecting chaos at the network level (e.g., iptables) tests infrastructure but misses agent-specific semantic failures. ChaosAgent intercepts tool invocations in Python user space, corrupting JSON schemas, injecting contradictory context, and simulating rate-limit backpressure to test LLM recovery reasoning.
Why automated Recovery Time Objective (RTO) scoring?▼
Teams need to quantify how quickly an agent detects a failed tool and routes to an alternate execution path. ChaosAgent computes a Resilience Index based on recovery steps, retry token waste, and state divergence.
Distributed System Architecture
Explore the multi-tier topology below. Switch between the interactive blueprint canvas, standard Mermaid.js flowcharts, and the step-by-step request simulator.
Chaos engineering framework for multi-agent LLM systems with configurable fault schedules, Byzantine interceptors, and automated recovery telemetry.
Engineering Post-Mortem & Next Iteration
I would add automated fuzzing against prompt injection safety filters to benchmark adversarial jailbreak resilience during chaos runs.
Next Case Study
06 // NEXT
BugPulse
AI-Powered Autonomous Bug Triage Platform