The End of the Frozen Chat: Hermes Agent Moves to Asynchronous Delegation

The era of waiting for an agent to finish its work is ending. Nous Research has updated Hermes Agent to allow its delegate tool to run subagents asynchronously, ensuring that delegated work no longer blocks the parent chat.

Previously, the delegate_task tool operated synchronously. When a parent agent spawned child agents—known as subagents—to distribute work, the parent agent would block during the tool call until every child completed its task. This design effectively froze the user's chat, preventing users from starting long-running tasks while continuing to work or steering an active run mid-flight.

The new update, introduced via issue #5586, implements a non-blocking path through an async_delegation toolset. Instead of waiting for completion, the system now spawns a background agent and returns a task_id immediately.

This shift changes how we manage agentic workflows in three specific ways:

First, it enables active supervision. The new toolkit allows users to monitor and manipulate running tasks through tools like check_task for status updates, steer_task to inject messages into a running task, and cancel_task to stop a process entirely.

Second, it preserves context efficiency. Subagents operate in isolation with their own conversation, terminal session, and toolset. Because only the final summary returns to the parent, the parent’s context window remains small, shielded from the child's intermediate reasoning or tool calls.

Third, it optimizes resource management. While subagents inherit the parent’s API key and credential pool—enabling features like key rotation on rate limits—users can route subagents to a cheaper model through config.yaml.

The architecture relies on background agents running as in-process threads, reusing the same machinery and credentials as the original synchronous method. However, concurrency remains managed; a synchronous batch still spawns children in parallel but is capped by delegation.max_concurrent_children, which defaults to 3.

This update transforms Hermes Agent from a sequential processor into a multi-threaded orchestrator. For developers building complex agentic loops, the ability to trigger background work without losing control of the primary interface is the difference between a prototype and a functional tool.

If you are running Hermes Agent, you can enable these features by running hermes update.

How will your workflow change once your primary agent no longer requires you to wait for its subordinates?