The Hermes subagent model is whatever your parent agent runs on, unless you change it. Set delegation.model and delegation.provider in config.yaml and every subagent switches to that model. Want a different model per subagent? That takes the official advanced-delegation plugin and its model routes.
📺 Watch: NEW Hermes Update Lets You Control Subagents Live
🔥 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
So there are two levels. One global worker model, built in. Or a menu of models the agent picks from per task. Here's how each works, how to choose models for fast, deep and local jobs, and where the money actually goes.
Where the cost of subagents really goes
Start with the reason this matters. The Hermes delegation docs make a sharp point: breaking a problem into well-specified subtasks takes frontier-level judgment, but carrying out a subtask that already has a clear goal, full context and an output contract usually doesn't.
Then the kicker. The docs say a parallel batch of subagents typically burns the large majority of a run's total tokens. The workers are where the volume is, so the worker model is where the cost lives.
That gives you the Nous-recommended pattern: keep your main session on a frontier model for planning, and pin subagents to an inexpensive model for the doing. Planning quality stays high, and spend drops where the tokens pile up.
I run Hermes daily inside my Agent OS with a profile per model, so I think about this split all the time. If you want my setup, the AI Profit Boardroom ($69/mo) gives you the Agent OS, a 30-day AI agent roadmap, daily tutorials as new tools drop, 4 coaching calls a week and the prompt library. 3,400+ members are inside.
📺 Watch: How to Rank #1 with Hermes AI SEO Subagents!
How to set the Hermes subagent model (built in)
No plugin needed for this. Open ~/.hermes/config.yaml and add two keys under the delegation block:
- delegation.model: the model subagents should use. The docs example is google/gemini-3-flash-preview.
- delegation.provider: optional. Routes subagents to a different provider, for example openrouter.
Leave both empty and subagents use the same provider and model as the parent. You can also set them from the terminal with hermes config set, or in Desktop under Settings, then Model, then Subagents. The Desktop page only affects subagents launched after you click Apply. It never changes your main model or a child that's already running.
Running subagents on a local model
For a local server, use the direct endpoint keys instead: delegation.base_url, delegation.api_key and delegation.model. The docs example points at http://localhost:1234/v1 with model qwen2.5-coder. If you leave out api_key, Hermes falls back to OPENAI_API_KEY only. There's also delegation.api_mode for endpoints whose wire format Hermes can't detect: chat_completions, codex_responses or anthropic_messages.
The resolution order
- delegation.base_url wins if it's set.
- Otherwise delegation.provider, with its full credential bundle.
- Otherwise children inherit the parent's provider and credentials.
- delegation.model applies in every case. Empty means the parent's model.
Worth knowing: delegation.fallback_providers gives workers their own fallback chain. Set it to an empty list and child fallback is switched off.
📺 Watch: New Hermes Subagents Update is INSANE!
The catch: one model for every subagent
The built-in pin is global. The docs say it straight: with the default strategy, delegate_task has no per-task model parameter, so every child in a batch runs on the same delegation model.
That's fine when all your workers do similar jobs. It hurts when one batch mixes a quick summary with a nasty bug hunt. The docs give three ways out: leave delegation.model unset for that session, use a delegation strategy plugin, or send the task to the kanban board, which does support a per-task model.
Need a second opinion on which models fit your workflows? Book a free AI strategy session.
Per-subagent models with model routes
The strategy plugin is advanced-delegation, from Nous Research's official plugins repo. It needs Hermes 0.22.5 or newer. Its model_map mode lets you define named routes, and the agent picks one per task. The full install and setup walkthrough is in Hermes advanced delegation. Here's the model side.
Each route has a name, a one-line description, a model, and a provider or base_url. Optional keys are reasoning_effort, service_tier and fallback_providers. The agent only ever sees the name and description, never the model ID or key. So the description is your real lever. Write it like a job ad: what this route is good at.
Two rules from the README. While model_map is active, delegation.model and delegation.provider are ignored, because the routes decide. And a task that picks no route uses the route you set as default, or the parent's model if there isn't one.
Every routed result reports what actually ran: route name, provider, model, reasoning effort and speed tier. Handy for checking the agent picks routes the way you expect. If it keeps sending hard jobs to the fast route, tighten the descriptions. Make the deep one say exactly which kinds of work belong there, and make the fast one say what it should never touch.
Want help writing routes that send the right work to the right model? Book a free AI strategy session and we'll set them up around your actual tasks.
Picking models for fast, deep and local jobs
The plugin docs use three example routes. They're a good template for thinking about any model choice.
| Route | Docs description | Docs example model | Pick a model that is | Cost and speed thinking |
|---|---|---|---|---|
| fast | Cheap and quick: lookups, summaries, simple edits | google/gemini-3-flash-preview via openrouter | Small, quick, good at following clear instructions | High volume, so low cost per token matters most |
| deep | Slow and careful: hard debugging and design work | claude-opus-5 via anthropic, reasoning_effort high | Your strongest reasoning model | Pricier per task, but a wrong answer costs more |
| local | Private: runs on this machine | qwen3-coder at http://localhost:8080/v1 | One your hardware runs well | No per-token bill, speed set by your machine |
Fast jobs
Lookups, summaries, formatting, simple edits. The subagent already has a clear goal, so you don't need a genius. You need cheap and quick, because there may be ten of these at once. Make this your default route so unrouted tasks land here.
Deep jobs
Hard debugging, architecture, anything where a wrong answer means redoing it. Here a stronger model is the cheaper choice overall. The deep route in the docs adds reasoning_effort high and an optional service_tier. Hermes checks the tier against the route, and if the route doesn't offer it, the spawn is refused instead of quietly running at another tier.
Local jobs
Private data, or bulk work you don't want on an API bill. Point the route's base_url at your local server. The trade-off is speed and capability, both set by your hardware and the model size you can run.
Don't guess on quality. Goldie Bench tests models on real agent tasks, which is the fastest way to shortlist what goes on each route. Pair the right model with the right skills, too: my list of the best Hermes Agent skills covers that half.
Prefer whole profiles?
The plugin's other mode, profile_routes, runs a subagent as one of your Hermes profiles, with its own model, credentials, skills and SOUL.md. If you already keep a profile per model like I do, that might be the simpler route. More on how spawning works in Hermes Agent subagents, and the Agent OS guide shows how profiles fit together.
FAQ
What model do Hermes subagents use by default?
The parent's model and provider. They only change when you set delegation.model, delegation.provider or delegation.base_url, or switch on a plugin strategy.
Can each subagent use a different model?
Not with the built-in settings. The pin applies to every child. Install the advanced-delegation plugin and use model_map routes, or use profile routes, to pick per task.
Does changing the subagent model affect my main agent?
No. The delegation keys only apply to children. Your main model stays as it is.
When do route changes take effect?
From the next session. The plugin builds the routing options when a session starts, so start a new chat with /new after editing.
Want your subagents on the right models from day one? Join the AI Profit Boardroom for the Agent OS and training, or grab a free AI strategy session and we'll pick your routes together.
