Can Hermes split a big job across helpers at once, and what does that cost?
Yes. Ask Hermes to split the work, for example “compare these five suppliers, one helper each”, and it can start up to 10 helper sub-agents at once with its delegate_task tool. Each helper works on its own part in a fresh conversation, and their results come back to your chat together when the last one finishes. Each helper is a separate model conversation on your model account, so the job usually finishes sooner but costs more than doing it in one chat, unless you point the helpers at a cheaper model.
This follows Hermes Agent v0.23.1, commit 3b42157, checked on 19 November 2026. Upstream calls the helpers subagents. We reviewed official documentation and source, including how Vavio sets up Hermes. We did not run a delegated job or measure its cost.
How do I get Hermes to use helpers?
Ask for it in plain words. delegate_task is one of Hermes’s default tools, and Hermes may also decide to use it by itself when a task splits naturally. The request works best when each part can be done without the others:
Check the pricing pages of these five suppliers. Use one helper per supplier, then give me one table with price, contract length and cancellation terms.
Hermes sends all the parts in one batch and gives each helper a goal and the context it writes for it. You can carry on chatting while they work. By default the results arrive as one message once every helper has finished, and Hermes then writes the combined answer.
If one part needs another part’s result, it isn’t a parallel job. Hermes should run the first part, wait, and then start the second with that result.
What does a helper know, and what can’t it do?
A helper knows only what Hermes writes into its task. It starts with an empty conversation: none of your chat history, earlier tool results or files you mentioned, unless Hermes passes them on. The one exception: when Hermes is working in a project folder, helpers also load that project’s instruction files, such as AGENTS.md. If a helper’s answer misses something you said earlier, the main agent didn’t pass it on; say it again and ask Hermes to include it.
Helpers get the same tools as the agent you’re talking to, never more. Five are always taken away:
| Helpers can’t | Why it matters to you |
|---|---|
Ask you a question (clarify) |
A helper guesses or reports back instead of stopping to ask |
Save to Hermes’s memory (memory) |
Only the main agent keeps facts for later chats |
Send messages to other chats or apps (send_message) |
A helper can’t message you or anyone else directly |
Create or change scheduled jobs (cronjob_manage) |
Recurring work stays with the main agent |
Start their own helpers (delegate_task) |
Helpers don’t multiply unless you raise delegation.max_spawn_depth |
Helpers can’t use the Kanban board either. Each one has its own terminal session, but they share the computer’s files: helpers editing the same folder can overwrite each other’s changes. For code in a git repository, delegation.worktree_isolation: true gives each helper its own copy; outside a git repository it quietly falls back to the shared folder. Helpers also inherit any computer-control tools, so on a computer with one desktop and browser, give screen or browser work to one helper at a time and use plain web tools for parallel research. We didn’t test several helpers using the screen at once.
Approvals. New Vavio computers start with Hermes’s approval prompts off, so helpers, like the main agent, run commands without asking you; Hermes’s built-in hard blocks still refuse the most destructive commands. If you turn approvals back on, where the request goes depends on where you’re chatting. In the hermes terminal chat, a helper can’t stop to ask, so Hermes refuses the risky command unless you set delegation.subagent_auto_approve: true. In the Hermes app or a messaging chat, the source routes the helper’s request to that chat, and it is refused if nobody answers. We haven’t tested either case. Read which actions Hermes asks you to approve before changing either setting.
How many run at once?
| Setting | Default in v0.23.1 |
|---|---|
Helpers in one batch, and background batches at once (delegation.max_concurrent_children) |
10; a new batch beyond that runs inside the reply, so the chat waits |
Turns each helper may take (delegation.max_iterations) |
250 |
Time limit per helper (delegation.child_timeout_seconds) |
None; a helper with no progress for 450 seconds (1,200 inside a tool) is interrupted and given 120 seconds to hand back partial work |
Levels of helpers (delegation.max_spawn_depth) |
1, so helpers can’t start their own |
Characters of each helper’s result passed back (delegation.max_summary_chars) |
Up to 24,000, less when the main agent’s context is nearly full; the rest is saved to a file |
A batch with more tasks than the limit is refused with an error, not cut short, so Hermes has to split it. There is no money cap. All helpers also share the computer’s memory and processor with everything else Hermes is doing.
What does it cost?
Each helper is its own model conversation, billed by the provider you configured for Hermes. Unless you change it, helpers use the same model, provider and credentials as your main agent. Every helper pays for its own instructions and tool definitions, its own reading and its own answer, and then the main agent pays again to read all the results and write the combined answer. Upstream’s guide notes that in a parallel batch the helpers typically use most of the run’s tokens.
Splitting saves time, not tokens. It can still make a long chat cheaper, because a helper doesn’t carry your whole conversation with it; see why a long conversation’s bill grows.
To cut the cost, keep your main agent on a strong model and give helpers a cheaper one:
delegation:
model: "your-cheaper-model"
provider: "openrouter" # optional; leave out to use the main agent's provider
In the Hermes app this is Settings → Model → Subagents in upstream’s Desktop guide; you can also ask Hermes to run hermes config set delegation.model <model>. The setting applies to every helper in a batch. With the default setup there is no per-task model choice, so a part that needs a stronger model means leaving the setting empty for that job, or using a Kanban card, which can name its own model. Check prices with your model provider.
What stops helpers, and what if Hermes restarts?
- Sending another message doesn’t stop them, and neither does the Stop button for the current reply. One exception: in a messaging chat such as Telegram where in-between messages are off, Hermes waits for its helpers inside the reply and sends one finished answer, and there a new message does stop them under the default busy-input setting (what happens when you message Hermes while it works).
/stopstops them, even after the turn has ended, and each stopped helper still reports back with its partial work, marked as interrupted. Hermes can also stop or redirect one helper on its own if you ask./newdepends on where you are. In a messaging chat or thehermesterminal chat, it stops them. In the Hermes app,/newjust opens a new chat and the helpers carry on; use/stopfirst, or delete the old chat.- Closing the Hermes app window or losing the connection doesn’t stop them, according to upstream’s guide.
- A Hermes restart does not resume a running helper. Its part is reported as
unknown, because Hermes can’t tell what it had already done. A helper that finished before the restart has its result delivered afterwards, within 48 hours, and occasionally twice.
Helpers are for work you’ll wait for in one chat. For work that must survive a restart, use a scheduled job or a Kanban card.
Helpers, /bg, /goal or a Kanban card?
| You want | Use |
|---|---|
| Several independent parts of one request done at the same time | Helpers (delegate_task) |
| One separate task without tying up the chat | /bg <prompt> |
| One agent to keep working on a task until it is done | /goal |
| A named agent to pick up work, on its own model, saved across restarts | Kanban cards |
| A recurring check without reasoning every time | A script-only check |
Does it work on a Vavio computer?
From the setup source, yes, but we haven’t run a delegated job on one. Vavio keeps Hermes’s default tools, including delegate_task, and doesn’t change any delegation setting, so the defaults above apply. The Vavio chat is Hermes’s own app; upstream’s Desktop guide shows running helpers above the message box with Steer and Stop controls, but we haven’t checked that view on a Vavio computer. Vavio keeps the computer and Hermes running, so a batch carries on when you close your laptop, though a restart still ends helpers that were mid-task.
Vavio provides a managed Hermes computer for $14.99 per month, with model usage billed separately by your provider. Helpers add model usage, not hosting cost. Compare the current hosting offer, or read how to run scheduled jobs on a cheaper model.
Sources
- Hermes v0.23.1 subagent delegation guide, checked 19 November 2026
- Delegation defaults (model, turns, concurrency, approvals), checked 19 November 2026
- Tools a helper can never use, checked 19 November 2026
- A batch larger than the limit is refused, checked 19 November 2026
- Helpers lost in a restart are reported as unknown, checked 19 November 2026
- Hermes v0.23.1 Desktop subagent settings, checked 19 November 2026
- Vavio hosting offer, checked 19 November 2026
Written by Vavio with AI assistance and checked against pinned official Hermes v0.23.1 documentation and source, and against Vavio's own computer setup source. This is a source review, not a live delegation trial or cost measurement.