Skip to content
vavio

Can my main Hermes agent and an extra agent hand work to each other?

Yes. Hermes keeps one Kanban board for every agent on a computer. Give your main agent a card to research something, then give your extra agent a second card to write it up, with the first card as its parent. The writing card waits until the research card is done, then starts in the extra agent with the research summary in its opening context. You don’t copy anything between chats. One Hermes gateway must be running to start the cards, and an agent in a normal chat can only create cards by itself after you turn on its kanban tools.

This follows Hermes Agent v0.23.1, commit 3b42157, checked on 18 November 2026. Upstream calls each agent a profile. Your main agent is the profile named default, and each extra agent has its own name. We reviewed official documentation and source, including how Vavio sets up extra agents. We did not run a handoff between two agents or measure its cost.

How do I hand work from one agent to another?

Create two cards and link them. From a terminal on the computer:

hermes kanban create "Research the three cheapest EU flights to Lisbon next weekend" --assignee default
# prints the new card's id, for example t_ab12

hermes kanban create "Write a short email to Sam recommending one of the flights" --assignee writer --parent t_ab12

From a chat, the same commands work as /kanban create .... Upstream documents /kanban in the terminal chat and on messaging channels such as Telegram, Discord and Slack. Then follow progress with hermes kanban list, hermes kanban show t_ab12 or the Kanban tab in the Hermes dashboard.

Step What happens
You create the research card for default It is ready, because it has no unfinished parent
You create the writing card for writer with --parent It waits in todo
Within about 60 seconds The dispatcher claims the research card and starts the main agent on it in a fresh session
The main agent finishes It records a summary, and the card moves to done
The parent is done The writing card becomes ready, and the extra agent starts with the research result

A card can have several parents, so a writer can wait for two researchers. The assignee must be the exact name of an existing agent. A card assigned to a name that doesn’t exist stays ready and never starts; hermes kanban show names the missing agent.

What does the second agent actually receive?

The writing card’s worker starts with a Parent task results section: each finished parent’s latest summary and its structured metadata, marked with how long ago it finished. It also sees its own card’s description and up to 30 recent comments.

Three limits matter:

  • Each summary or metadata field is cut at 4,096 characters. A long research write-up will be truncated. Upstream’s guide says the handoff arrives verbatim; the pinned source caps it. Ask the first agent to end with a short summary, and put the full material in a file.
  • Only declared files are kept. By default a card works in a scratch folder, which is deleted once that card and the cards waiting on it are finished. Files the first agent declares as artifacts when it completes are copied to the card’s saved attachments, and their paths reach the next card with the summary. For anything else, give the card a fixed folder with --workspace dir:<path> and name that folder in the next card’s description.
  • The result is a snapshot. It is the parent’s summary at the moment it finished, not a live link to the first agent’s memory or conversation.

Can one agent assign work to the other by itself?

Only after you allow it. The kanban_* tools are off by default in chats. A default or “all tools” setting is not enough; enable them for that agent and channel, then start a new chat:

hermes -p default tools enable kanban
hermes -p default tools enable kanban --platform telegram

After that, you can ask your main agent “research this, then have the writer draft the email”, and it can call kanban_create with the writer as assignee and its own card as parent. Workers started from the board already get the tools they need to report back, add comments and create follow-up cards.

What has to be running?

One Hermes gateway that runs the dispatcher. The dispatcher runs inside the gateway by default and checks the board every 60 seconds. Only one gateway on a computer dispatches at a time; it holds a lock file, and any others skip. To run a card, the dispatcher starts a small worker client for the assigned agent, which asks the computer’s gateway to run the turn as that agent and starts the gateway if needed. So the extra agent doesn’t need its own chat open, or its own gateway already running, for its card to start.

If no gateway is dispatching, cards stay ready and nothing happens. hermes kanban create warns you when that is the case. When a card is created from a messaging chat such as Telegram, that chat is subscribed to it, and you get a message with the first line of the summary when it finishes, is blocked or fails. A card created in the terminal isn’t subscribed to any chat.

What does a handoff cost?

Each card runs as a fresh session on the assigned agent’s model, unless the card names its own model with --model. So the research costs whatever your main agent’s model charges for that work, and the writing costs whatever the extra agent’s model charges. A fresh session doesn’t carry a long conversation with it, which can make a card cheaper than doing the same work at the end of a long chat. Check prices with your model provider.

These defaults bound it:

Setting Default in v0.23.1
Cards running at once on the computer (kanban.max_in_progress) Total memory ÷ 512 MiB, kept between 2 and 8 (on a 8 GB computer, 8)
Failed starts or crashes before a card is blocked (kanban.failure_limit) 2
Turns per card The agent’s usual turn limit; --goal cards stop after 20 judged turns
Time limit per card None unless you set --max-runtime

Upstream’s guide calls the concurrency limit unlimited when unset. The pinned source derives it from memory on Linux. There is no money cap on the board.

Is this the same as /goal, /bg or delegation?

You want Use How it runs
One agent to keep going on one task in this chat /goal Turn after turn in your conversation, checked by a judge
One separate task, same agent, without tying up the chat /bg <prompt> Once, in its own background session
Quick helpers for the agent you’re talking to Delegation (delegate_task) Helper sub-agents that run in the background and report back to the same chat; stopped by /stop, /new or Hermes exiting
A named agent to take over after another finishes Kanban cards with --parent Saved on the board; each card starts in the assigned agent when its parents are done

The board is the one option where the work is saved. Cards are rows in a database file, so a gateway restart doesn’t lose them. A worker that dies is marked crashed and the card goes back to ready for a fresh attempt. --goal on a card combines the two: the card’s worker runs the /goal loop, using the card’s title and description as what counts as done.

Does it work on a Vavio computer?

From the setup source, nothing should stop it, but we haven’t tried a handoff on one. On a Vavio computer, an extra agent is a standard Hermes profile under the same Hermes folder as the main agent, run by the same system user, with its own model settings. Hermes deliberately places the board in that shared folder, so both agents see the same board. One gateway serves every agent and is kept running, so cards can start while your laptop is off. Vavio doesn’t disable the Kanban commands, tools or dashboard tab, and it doesn’t turn on the kanban tools either; that stays your choice per agent. We haven’t checked whether the Vavio web chat runs /kanban the way upstream’s chats do. If it doesn’t, send the same /kanban commands from a messaging chat such as Telegram, if you have connected one to that agent. Vavio’s MCP server and CLI don’t read or change the board.

Two agents working at once use the computer’s memory and processor together. Each card also bills its own model separately from hosting. Vavio provides a managed Hermes computer for $14.99 per month, with model usage billed separately. If you want each agent’s data fully separate, read what two profiles share first, and compare the current hosting offer.

Sources

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 two-agent trial.

More from the Vavio blog: Hermes Agent guides