On July 13, a NousResearch/hermes-agent commit added durable background-completion delivery, while an openclaw/openclaw commit added a live-turn hold for native background work. A background agent can finish correctly and still fail its user: the process may restart before the completion is delivered, or the interface may declare a turn done while the work that will produce its real answer is still running.
Hermes' delegation change writes a completion record before queueing it. On a fresh process, pending records can be restored and are only marked delivered after the consumer acknowledges injection. That is a receipt: the task result is not treated as delivered merely because the worker stopped.
Open turns are a delivery boundary too
OpenClaw's change addresses a different gap. Its Claude CLI runner watches the CLI's authoritative background-task list. If a native subagent or workflow remains, an interim success result does not close the live turn; the runner waits for the post-drain result and preserves the earlier text. Its new tests intentionally exclude a background shell task, because that kind is killed at exit and does not hold the final answer.
The useful question is not “did it run?”
The common lesson is narrower than a claim of durable agents: systems need an explicit handoff between completed work and a reader-visible result. Hermes makes that handoff restart-aware; OpenClaw makes it turn-aware. Neither commit measures end-to-end reliability, but both make the next test clear: interrupt the process, close the interface, or let a subtask finish late—and see whether the result still has a named route back.