Saboteur: Chaos Monkey for AI agents

AI agents look great in demos and fall apart in production. APIs return 500s, rate limits kick in, tool calls come back malformed and context gets lost. Most agents don't handle this well: they loop on retries, invent results, or just give up. There was no standard way to test for this before shipping, so I built one. Saboteur runs your agent while deliberately breaking things around it. You pick a chaos profile (a fixed set of faults with a seeded RNG, so runs are repeatable) and Saboteur launches a cohort of N agents on the same task at once. Every fault, tool call and recovery gets written to an event log and streamed live to a dashboard grid: green is healthy, amber is recovering, red is crashed, cyan passed. At the end you get a Resilience Scorecard with behavioral metrics (recovery time, waste factor, crash rate) and when a ground-truth check exists, verdict metrics like survival rate and deception detection. It works on agents you don't own too. Point your agent's OpenAI-compatible client at Saboteur's proxy and it injects faults into the wire traffic with zero code changes. There's also a stdio shim for MCP tool servers and a GitHub Action so agent resilience can be a required check on every PR. The demo cohort runs (N=50, hell_mode profile) were captured on an AMD MI300X running vLLM, comparing Llama 3.1 8B/70B and Gemma 4 model families. Local dev runs entirely offline against llama.cpp or a deterministic mock, no GPU or API key needed.