Opening the door to agent autonomy: The architecture behind Rovo’s agent harness

This post introduces the architecture of Rovo's new agent harness, detailing its split-plane design (data and control planes), isolated compute sandboxes, asynchronous multi-agent coordination via interactive sub-agents, and error recovery mechanisms (discovery probes, checkpointing). It also covers programmatic tool calling, dynamic model selection, dynamic/static tool loading, and a callback bridge for secure tool invocation. The system is integrated with an automated evaluation pipeline for continuous improvement.