Developer Builds 'Ballast' Layer on OpenTelemetry to Track AI Agent Reliability

A developer has released an open-source tool called Ballast that sits on top of OpenTelemetry to address a gap in AI agent observability. While OpenTelemetry's GenAI conventions capture operational metrics like token counts, latency, and error status, they do not indicate whether an agent's output was actually acceptable or safe. Ballast adds a reliability layer by introducing prompt contracts, guardrail verification, and bounded loop tracking as attributes on existing OpenTelemetry spans. The tool is designed to be narrow in scope — it wraps existing model calls without replacing runtimes, tracing backends, or orchestration frameworks. To prevent false confidence, every span includes a completeness flag signalling whether the instrumentation captured the full picture of what the agent did.