I reviewed Hermes Agent nine days ago. The one thing I flagged as the asterisk on subagents just got fixed.
Tonight, Nous Research shipped async subagents for Hermes Agent. The delegate_task tool used to block your chat while subagents ran. Now you can kick off a research task, keep working, check on it, steer it, and collect the results when ready. It changes how I run the agent.
The Fix in 30 Seconds
The old delegate_task was synchronous. You handed a subagent a goal and waited. The whole chat session blocked until the child returned a result. It was fine for quick lookups. Painful for real research.
The new async_delegation toolset has six tools: delegate_task_async, check_task, collect_task, steer_task, cancel_task, and list_tasks. delegate_task_async returns a task_id immediately. Your agent keeps working. You check on the subagent when you want, steer it if it goes off-track, collect the results when ready.
The Fan Out, Keep Working Pattern
Here is the worked example that matters. Start a research subagent on topic A. While it runs, draft the outline for topic B. Check on A every few minutes. If A goes off-track, steer it with a correction. When A finishes, collect the full results and merge them into B’s outline. The whole time, your chat session stays alive and productive.
That is exactly the flow I run every day on TRT. I kick off a research pass on one topic while I draft the lede for another. The difference is that until tonight, I had to either wait for the research to finish or run it manually in a separate session. Now the agent does both in one session.
Why This Is the Asterisk on My Jun 6 Review
In my Hermes Agent review, I called out the closed learning loop and multi-agent Kanban as the biggest strength. Subagent fan-out was a killer feature with one asterisk: the chat blocked while subagents ran. I wrote that subagent delegation was “the most underappreciated feature in the toolkit” but noted the blocking behavior as a real operational cost.
Tonight that asterisk is gone. The feature is still the same closed learning loop. The Kanban is still there. But now the loop doesn’t stop when you delegate.
The Persistent-Agent Cross-Cut
OpenAI just bought Ona for persistent cloud agents. Hermes Agent just shipped the in-process version of the same idea tonight. OpenAI is buying the persistent-agent future. Nous Research is shipping it for free in MIT-licensed code you can run on your own box.
Hermes Agent runs across every surface I use: desktop, iMessage via Photon Spectrum, and the Obsidian + NotebookLM memory stack. Async subagents make all of those more useful because a research subagent can finish in the background while the parent agent keeps working on whatever surface you are on.
How to Access It
Run hermes update. The new async_delegation toolset is opt-in — the old delegate_task is still there unchanged. The feature shipped as part of the ongoing v0.16.x series (tracked since GitHub issue #5586 on Apr 6, community PRs merged this week).
The Honest Read
The new toolset is for long subagent work. If your subagent tasks finish in 30 seconds, the old synchronous delegate_task is fine. Async is the right answer when you are kicking off research, big refactors, multi-file audits, or anything that takes more than a coffee to finish.
One honest caveat: the child agents run as in-process threads. The chat does not block, but your machine’s resources do. Long-running subagents still consume memory and CPU. The architecture is correct for a desktop operator tool. It is not a production batch scheduler.
If you read my review and shrugged off the subagent feature because the chat blocked, go update and try it again. It is a different experience now.
News analysis based on Nous Research’s announcement and the Hermes Agent GitHub repo. I’m running the same Hermes Agent I reviewed on Jun 6. No sponsorship, no business relationship with Nous Research.
