Hermes delegation is how Hermes Agent hands work to subagents: the delegate_task tool spawns child agents that each get a fresh conversation, work on their own task, and send back only a final summary. You control how many run at once, how deep they can nest, and which model or profile each one uses, and since v0.22.6 the /delegation command saves your choice to config.yaml for the next session.
📺 Watch: New Hermes Delegation Update: AI Agent Teams!
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
Hermes Agent is built by Nous Research. This page is based on the official Subagent Delegation docs, the configuration reference and the v0.22.6 release notes published on GitHub on 2 November 2026.
Our members share their delegation setups and model routes inside the AI Profit Boardroom ($69/mo). Want help deciding which jobs to hand to agents first? Book a free AI strategy session.
How Hermes delegation works
The delegate_task tool spawns child agents with isolated context, inherited tool access and their own terminal sessions. Each child works on its own. Only its final summary goes back into the parent's context.
That is the whole point. Ten file reads and forty tool calls happen inside the child. The parent sees one short report: what was done, what was found, which files changed and any problems.
Top-level delegation runs in the background. Hermes returns a handle straight away so your conversation keeps going, then posts the result back as a new message when the child finishes.
You can delegate one task or a batch:
- Single task: one goal plus context, for example debugging a failing test.
- Parallel batch: a list of tasks that run at the same time, for example researching three topics at once.
Each task can also carry an output_schema, so the child must return structured data, and up to 8 images for visual jobs like comparing a page against a design mock.
For more on what subagents are good for, read our Hermes Agent subagents guide.
Hermes delegation context: subagents know nothing
The docs put this in a warning box. Subagents start with a completely fresh conversation. They have zero knowledge of your chat history or earlier tool calls. Their only context is the goal and context fields the parent fills in.
One exception: when the parent has a workspace folder, each child gets that project's context files, such as AGENTS.md. SOUL.md is left out.
So a vague goal like "fix the error" fails. A good goal names the file, the line, the error text and the project path. If your delegated tasks come back wrong, the brief is usually the problem, not the model.
📺 Watch: Hermes Agent Just Got Smarter at Managing AI Teams!
Hermes delegation limits: width and depth
Two settings in config.yaml control how big a delegation tree can get.
delegation.max_concurrent_children caps how many subagents run in parallel per batch. The configuration reference sets the default at 3, with a floor of 1 and no ceiling. You can also set it with the DELEGATION_MAX_CONCURRENT_CHILDREN environment variable. If the model sends more tasks than the cap, delegate_task returns an error explaining the limit. It does not quietly drop tasks.
delegation.max_spawn_depth sets how deep the tree goes, clamped between 1 and 3:
- 1 (default): flat. Children cannot spawn their own children.
- 2: a child started as an orchestrator can spawn leaf grandchildren.
- 3: three levels.
The agent opts into orchestration per call with the orchestrator role. delegation.orchestrator_enabled set to false forces every child back to a leaf.
Cost multiplies fast. The docs point out that depth 3 with 3 children per batch can reach 27 leaf agents running at once. One-shot runs, like hermes chat -q, have their own cap, delegation.oneshot_max_children, which defaults to 2.
📺 Watch: NEW Hermes Kanban Agent is a Game CHANGER!
Hermes delegation strategies
By default every child inherits the parent's provider and model. You can change that two ways.
The simple way is delegation.provider and delegation.model. Set them and every child runs on that model. The docs give the classic use: a cheap, fast model for narrow subtasks while your main agent runs a stronger reasoning model. The catch is that this pin is global, so every child in a batch gets the same model.
The flexible way is delegation.backend, which picks a delegation strategy:
| Backend | How each child is routed | Where it comes from |
|---|---|---|
| default | Uses delegation.provider and delegation.model, or the parent's route when they are empty | Built in |
| model_map | You list named options such as fast and deep, each with a description and a provider and model; the agent picks one per task | Official advanced-delegation plugin |
| profile_routes | Each child runs as one of your Hermes profiles, with that profile's model, credentials, skills and persona | Official advanced-delegation plugin |
A few rules from the docs:
- Each strategy's settings live under delegation.backends followed by the backend name.
- While model_map or profile_routes is active, delegation.model and delegation.provider are ignored. The routes decide.
- The model only sees the option names and descriptions you wrote, never raw provider or key fields.
- If the chosen backend's plugin is missing or disabled, delegate_task fails with an error naming the fix. It never silently runs children on another route.
- Profile routes share that profile's credentials and terminal policy, so only list profiles you trust with that access.
The plugin also adds hermes delegation config for creating and editing routes. We go deep on the plugin, model maps and profile routes in our Hermes advanced delegation guide.
The /delegation command in Hermes delegation
The v0.22.6 release notes list "/delegation choices saved for the next session" among the changes since v0.22.5. The /delegation slash command comes from the advanced-delegation plugin.
Here is how it behaves, per the docs:
- /delegation with a backend name, for example /delegation default, saves delegation.backend to config.yaml.
- The new choice applies from the next session, after /new. A running chat keeps the backend it started with, because the task options the model sees are part of its cached tool list.
- Add --now to switch the current chat as well. Hermes rebuilds the chat's tools at its next turn, at the cost of one prompt-cache miss.
- Editing delegation.backend by hand behaves the same as the command.
- Without the plugin, hermes config set delegation.backend default puts you back on the default route.
If a task sends a parameter the chat's backend does not take, Hermes refuses it with a message naming the backend, the parameter and the fix. It does not fall back to a default route.
To get the update: CLI users run hermes update. The Microsoft Store app updates through the Store, and Docker and Hermes Cloud use the stable or latest image. The release notes say the Desktop app from the .exe or .dmg installer gets it with the next bundled release.
Hermes delegation tips
- Write full briefs. Put file paths, error text and the goal in every task.
- Start flat. Keep max_spawn_depth at 1 until a job needs an orchestrator.
- Use cheap models for narrow jobs. Lookups and summaries do not need your strongest model.
- Use output_schema when you plan to read the result in code or a spreadsheet.
- Watch the work. In the TUI, Ctrl+T opens the live monitor for running subagents.
Want a plan for which tasks to delegate across your business? Book a free AI strategy session. For how delegation fits with skills, memory and routines, read the Agent OS guide. To choose models for your routes, Julian keeps his own model tests on Goldie Bench.
Hermes delegation FAQ
Do subagents see my chat history?
No. Each subagent starts with a fresh conversation. It only knows the goal and context the parent passes, plus project context files when a workspace is set.
How many subagents run at once by default?
The configuration reference sets delegation.max_concurrent_children to 3 by default. You can raise it, with no hard ceiling.
Why did /delegation not change my current chat?
The choice is saved to config.yaml and applies from the next session. Add --now to switch the current chat at its next turn.
Do I need a plugin for Hermes delegation?
No. delegate_task and the default backend are built in. The model_map and profile_routes strategies and the /delegation command come from the official advanced-delegation plugin.
Want ready-made delegation setups and weekly help? Join the AI Profit Boardroom. Want a one-to-one plan? Book a free AI strategy session.
