runtime-safety

Runtime safety net for LLM agents that detects runaway token growth, classifies failure patterns, and suggests fixes without extra LLM calls. The project includes a Rust core, Python SDK, examples, benchmarks, and benchmark claims across multiple models and runs.

A research article describing an empirical runtime-monitoring approach for multi-turn LLM agents, backed by 3,175 runs across four benchmarks. It also points to an open-source implementation, state-harness, with Rust/Python support, LangGraph and CrewAI adapters, CLI tooling, and OpenTelemetry export.