Open-Source Background Agents Brings Four-Layer Fault Tolerance to Long-Running AI Tasks

Background Agents, an open-source project built on the Open-Inspect framework, is designed to handle multi-hour autonomous coding sessions that typical agent demos cannot sustain. It implements a four-layer fault-tolerance system that escalates failures through limited retries, fallback model hand-off, parallel evaluator shadow monitoring, and feedback-injected reruns. The project runs tasks inside isolated Docker sandboxes, allowing parallel sub-tasks with independent file systems and git directories, while a control plane manages sandbox provisioning and graceful shutdown on failure. The architecture is based on an approach Ramp described publicly, but this open-source release exposes the orchestration logic that most agent frameworks leave undocumented. The project has reached 3,057 GitHub stars and is currently trending fourth among TypeScript repositories.