BalaganAgent brings chaos engineering to production AI agents. Wrap your agent and inject controlled failures: tool timeouts, rate limits, malformed outputs, latency spikes, hallucinated values, context corruption and budget exhaustion - then measure MTTR and an SRE-style reliability score. Get terminal/JSON/Markdown/HTML reports. Wrappers for CrewAI, Claude Agent SDK and example wrappers for AutoGen, LangChain. Open source: pip install balagan-agent

Humans in the Loop — A free community for people who love using AI to ship faster

A free community for people who love using AI to ship faster

I kept seeing the same pattern: agents look great in demos and even pass evals, then production hits tool calls time out, rate limits kick in, outputs come back malformed, latency spikes, context gets corrupted, and the agent fails silently. Balagan Agent is my attempt to bring SRE-style chaos engineering to agents: inject controlled faults before users do, then measure how fast (and how well) the agent recovers (MTTR + reliability score). It started as a tiny fault-injection layer, and evolved into chaos levels + metrics + reports, plus wrappers for popular agent frameworks. If you’re building agents you actually want to run in prod, I’d love feedback on: which failure modes hurt you most, and which integrations you want next.