The video explains that sub‑agents are lightweight processes running on target machines, managed by the Harness primary agent. They execute tasks, collect artifacts, and monitor health, communicating via secure gRPC. Sub‑agents register with the primary agent at startup, receive job payloads, and report status. Features include parallelism, dynamic scaling, and containerized deployment, which isolate failures, improve resource use, and simplify upgrades. Configuration uses YAML for resource limits and retry policies. Security relies on mutual TLS and token‑based authentication. Logging integrates with Harness observability, and sub‑agents can be versioned independently for backward‑compatible releases.