Hermes delegation lets one Hermes Agent hand work to subagents: separate child agents that each get a fresh conversation, do one job and send back a summary. You can run several at once, send cheap tasks to a fast model and hard ones to a strong model, and redirect a child while it works. The stable v0.22.6 release on 2 November 2026 also made your /delegation choice save for the next session. Below is how delegation works, the settings that matter, model routing, and how to watch and steer subagents. My video is right under this intro.
Key takeaways
- Hermes delegates with the
delegate_tasktool: one task or a parallel batch. - Each subagent gets its own fresh context; only its final summary comes back.
- Route subagents to a different model with
delegation.model, or named routes with the advanced-delegation plugin. /delegationchoices now save for the next session; add--nowto switch the current chat.- Want a team of AI agents doing the work? Claim a free strategy session.
How Hermes delegation works
From the Subagent Delegation docs: delegate_task spawns child agents with isolated context, the parent’s tool access and their own terminal sessions. Each child works independently, and only its final summary enters the parent’s context, which keeps your main chat clean.
Delegation runs in the background. Hermes hands back a handle straight away so you can keep talking, then posts the result as a new message when the child finishes.
delegate_task(tasks=[
{"goal": "Research topic A", "context": "Focus on recent primary sources"},
{"goal": "Fix the build", "context": "Project root: /home/user/project"}
])The rule that matters most: a subagent knows nothing about your chat. Give it the full context in the task, not “fix the error”.
The delegation settings that matter
All in the delegation: block of config.yaml (configuration docs):
| Setting | Default | What it does |
|---|---|---|
max_concurrent_children |
3 | Subagents running in parallel per batch |
max_spawn_depth |
1 | 1 = children can’t delegate; 2 or 3 allows nested teams |
orchestrator_enabled |
true | Kill switch for orchestrator children |
oneshot_max_children |
2 | Cap for one-shot runs (hermes chat -q) |
worktree_isolation |
false | Give each child its own git worktree |
gateway_mode |
auto | How chat apps wait for subagent results |
If the agent asks for more parallel tasks than the cap, delegate_task returns an error explaining the limit instead of quietly dropping tasks.
Want a team of AI agents running parts of your business? Inside the AI Profit Boardroom you get the Agent OS — Hermes, OpenClaw, OpenCode, Claude and Codex in one place — plus step-by-step Hermes training and four live coaching calls a week.
Want a plan for your business first? Claim your free strategy session.
Join the AI Profit Boardroom →Claim Your Free SEO Strategy Session →
Send subagents to a cheaper model
By default subagents use the parent’s model. A common setup from the docs is a strong planner with inexpensive workers:
delegation:
provider: openrouter
model: google/gemini-3-flash-previewSet just model to switch models on the same provider, or base_url and api_key to send subagents to any OpenAI-compatible endpoint. Workers can have their own fallback chain with delegation.fallback_providers.
Model routes and the /delegation command
The official advanced-delegation plugin adds two smarter strategies, set with delegation.backend:
- model_map: you list named options (for example
fastanddeep), each with a description and a model, and the agent picks one per task. - profile_routes: each subagent runs as one of your Hermes profiles, with that profile’s model, skills and persona.
The agent only sees the names and descriptions you wrote, never keys or raw model fields, and each result says which route it actually ran on. The plugin also adds hermes delegation config to create and edit routes.
What changed: since the 23 October change, now in stable v0.22.6, /delegation <backend> saves to config.yaml and applies from your next session (/new). Add --now to rebuild the current chat’s tools at its next turn, at the cost of one prompt-cache miss. /delegation default switches back.
Watch and steer subagents live
- The CLI, TUI and Desktop show running subagents above the message box, with task names, time and latest activity.
- Ctrl+T opens the live roster in the CLI and TUI; in Desktop, expand Subagents and pick a worker.
- Steer sends a correction (“focus on pricing instead”) without stopping the child. Stop ends just that worker.
- On Telegram, Discord and Slack,
/agentslists background delegations with live activity per child.
Steering is queued, not instant: the child reads it at its next checkpoint, and if it finished first, Hermes tells you the steer didn’t land.
Want this set up for you instead of doing it yourself? Claim your free strategy session →
Live transcripts and results
Every child writes a live log under ~/.hermes/cache/delegation/live/<id>/, with its text, tool calls and results, and a manifest.json for the batch. You can tail -f a log while it runs. Logs older than 7 days are cleaned up automatically. Need structured data back? delegate_task accepts an output_schema.
My verdict
Start with plain delegate_task batches and a cheaper worker model. When you know which jobs need a strong model, add a model_map with fast and deep routes and switch it on with /delegation model_map --now. That’s the practical path to Hermes delegation doing real work for you.
About Julian Goldie
I’m Julian Goldie — founder of Goldie Agency and the AI Profit Boardroom. I test new AI agents almost every day for 432K+ YouTube subscribers and run Hermes, OpenClaw, OpenCode, Claude and Codex side by side in my own Agent OS. Claim your free strategy session or join the AI Profit Boardroom.
Want another angle? Read the AI Profit Boardroom guide to Hermes Delegation.
FAQ: hermes delegation
What is Hermes delegation?
It is how a Hermes Agent hands tasks to subagents with the delegate_task tool. Each subagent gets fresh context, works independently and returns only a summary to the main chat.
How many subagents can Hermes run at once?
The config default for max_concurrent_children is 3 per batch, with no hard ceiling. Raise it in the delegation block of config.yaml.
Can Hermes subagents use a different model?
Yes. Set delegation.provider and delegation.model, or use the advanced-delegation plugin’s model_map to give each task a named route such as fast or deep.
What does the /delegation command do?
It comes with the advanced-delegation plugin and switches the delegation backend. Since v0.22.6 it saves to config.yaml for the next session, and –now applies it to the current chat.
Can I steer a subagent while it runs?
Yes. Use Steer in Desktop, the live roster in the CLI or TUI, or have the parent agent send a steer action. The child reads it at its next checkpoint.
Can Hermes subagents spawn their own subagents?
Only if you raise max_spawn_depth to 2 or 3. At the default of 1, children cannot delegate further.
Ready to put Hermes to work in your business? Join the AI Profit Boardroom for the Agent OS, setup training and daily updates.
Or claim your free strategy session and we’ll map out your setup.
Starting from zero? Get the free AI course and community.
Join the AI Profit Boardroom →Claim Your Free SEO Strategy Session →Get the FREE AI Course + Community →
Related reading
Last updated November 2026. This page is a living guide to hermes delegation — the facts here move fast and it is updated as they do.
