In a September 7 NousResearch/hermes-agent commit, a child must now wait, kill, or formally hand off background work before it returns. That matters when a parent hears “the CI watcher is running,” waits for a result, and never gets one because child teardown killed the watcher: a result needs an owner who can actually receive it.
The new handoff action lets a child transfer a running process it owns to its parent. Hermes changes the process’s owner and session under a registry lock, carries a one-sentence purpose with its completion, and lets the parent poll, stop, or receive the result. The transfer is deliberately narrow: it is only for a live child-owned process, rejects foreign or exited work, and caps each child at three handoffs.
“Running” is not a status if nobody will hear the finish
That boundary is the story. Hermes’ public delegation guide says that child-owned background processes are killed when the child ends unless the child waits, kills them, or hands them off. Sharing a terminal environment does not silently transfer ownership. The runtime now refuses to let a child’s final prose do that job by implication.
The focused handoff test makes the difference concrete. It hands one child-owned process to the parent and checks that its completion arrives with the purpose “CI watcher for PR 1.” A sibling process that was not handed off is recorded as orphaned before cleanup terminates it. The parent gets a receipt from the runtime, not a hopeful sentence from the child.
Handoff is not durability
The tempting headline would be that Hermes made child work survive. Its own source says otherwise. Handoff moves a live process to a parent; it does not checkpoint it, make it restart-safe, or guarantee a remote CI service will stay reachable. For a watcher, the documentation still prefers a simpler pattern: return the PR number and SHA, then let the parent start the watcher it intends to own. The change is included in Hermes’ v2026.9.7 patch rollup, but that is packaging evidence, not a report of adoption or reliability in the wild.
That restraint is the transferable rule. In any nested agent system, a “done” message should mean one of three things: the work finished, ownership was explicitly transferred, or the work was stopped and disclosed. The next useful signal is a test of what happens when the parent is interrupted—because a watcher only becomes trustworthy when its ownership remains visible at the next failure boundary.