Field notes · AI agents, Hermes & Claude Code By Julian Goldie

Hermes Agent Kanban Orchestrator: Set Up A Planner Profile And Worker Board In 8 Steps (2026)

One profile plans, the rest execute. Here is the exact setup order, the config keys to check, and the mistakes that leave cards sitting in ready.

Get The AI Profit Stack Join AIPB →
  • 01 1,000+ done-for-you AI agent workflows
  • 02 5 live coaching calls / week with me
  • 03 7-day refund, cancel anytime
  • 04 3,400+ AI operators inside

A hermes agent kanban orchestrator is one Hermes profile whose only job is planning: it turns your goal into cards on the board, hands each card to a named worker profile, and stays out of the actual work.

Here is how I'd set it up, in eight steps, using only commands and config keys from the official Hermes Kanban reference.

The example running through every step is a small accounting practice that wants a monthly client newsletter researched, drafted and checked without anyone chasing the pieces.

If you want a second opinion on your own setup before you build, book a free SEO strategy session with my team.

Why The Planner Lane Comes Before Anything Else

Hermes Kanban is a durable board shared by every profile on the machine.

Each card is a row in ~/.hermes/kanban.db, and each worker the dispatcher spawns is a separate OS process with its own identity.

The docs are blunt about the planner's role: "A well-behaved orchestrator does not do the work itself."

It decomposes, links, assigns and steps back.

So the first design call is not which model or which board.

It is deciding which profile plans and which profiles execute, and keeping that line hard.

The behaviour guidance (anti-temptation rules, a profile-discovery step and a decomposition playbook) is injected into the system prompt automatically, so you don't install anything extra for it.

Step 1: Build The Worker Roster First

The planner can only assign cards to profiles that already exist.

For the accounting practice, three workers cover the newsletter: researcher gathers the month's tax-deadline reminders, writer drafts the issue, and checker reviews it against the brief.

Create those profiles, then confirm them:

hermes profile list

This matters more than it looks.

The dispatcher fails quietly on assignee names it can't resolve: the card sits in ready and logs a skipped_nonspawnable event instead of running.

That is why the injected guidance tells the planner to look up real profile names before it writes a single card.

Step 2: Describe Each Profile So Routing Works

Descriptions are what the built-in decomposer reads when it picks an assignee.

You can write them yourself or let Hermes draft them from each profile's installed skills and model:

hermes profile describe researcher --text "Finds and summarises source material with links"
hermes profile describe writer --auto

Profiles without a description still appear in the roster and are routable by name, just less precisely.

Keep each description to what the profile does, not what it is called.

Step 3: Create The Planner And Switch On The Kanban Toolset

A normal chat profile has zero kanban_* tools.

The board view in Desktop does not grant them either; you enable the toolset per profile and per platform:

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

Each platform keeps its own selection under platform_toolsets.<platform> in config.yaml.

Start a fresh chat afterwards, because existing conversations keep the tool schema they started with.

Once enabled, the planner gets kanban_create, kanban_link, kanban_list, kanban_unblock and kanban_comment for steering other cards.

🔥 Want a planner profile and worker roster you can copy instead of building from zero? Inside the AI Profit Boardroom you get step-by-step video trainings, the installable Agent OS with Hermes ready to go, four coaching calls a week and 3,600+ members building agent setups like this one. → Join the AI Profit Boardroom

Step 4: Strip The Planner Down To Kanban, Gateway And Memory

The worker lanes page defines the orchestrator lane as a profile whose toolset includes kanban but leaves out terminal, file, code and web.

The reference goes one step further and pairs the planner with only kanban, gateway and memory.

With that set, the planner literally cannot draft the newsletter itself, even if the model is tempted.

That is the whole point: if it can't do the work, it has to delegate the work.

Check the result with hermes tools or the dashboard toolset checkboxes before moving on.

Step 5: Split Models Between Planner And Workers

The docs' cost strategy is a frontier model on the orchestrator profile and inexpensive models on the workers.

Every profile reads its own ~/.hermes/profiles/<name>/config.yaml, because the dispatcher injects a profile-scoped HERMES_HOME when it spawns a worker.

The reference uses placeholders such as your-frontier-model and your-inexpensive-model rather than naming models.

When one card needs more muscle, pin it without touching the profile:

hermes kanban set-model <task-id> <model> --provider <provider>
hermes kanban set-model <task-id> none

The second line clears the override again.

You can also pass --model and --provider at create time, or use the model dropdown in the dashboard.

Step 6: Initialise The Board And Start The Gateway

hermes kanban init
hermes gateway start

The dispatcher lives inside the gateway by default (kanban.dispatch_in_gateway: true) and ticks every 60 seconds (dispatch_interval_seconds).

No gateway means no dispatch: ready cards wait, and hermes kanban create warns you about it at creation time.

The old standalone hermes kanban daemon is deprecated, so don't build around it.

If the practice already runs its Hermes bot on Telegram, the same gateway is doing both jobs; my Hermes Agent Telegram setup covers that side.

Step 7: Hand The Planner Its First Goal

Open a new chat with the planner and give it the outcome, not the steps: "Produce the November client newsletter: deadline reminders, a short article on expense records, and a final check against our style notes."

A good planner turn looks like the docs' own example: two research cards in parallel, then a writing card with both as parents.

The child stays in todo and is promoted to ready only when every parent reaches done.

You can also add cards by hand when you know exactly what you want:

hermes kanban create "Draft November newsletter" --assignee writer --parent <research-task-id>

Write the shared decisions into each card body yourself if you go manual, because workers can't see sibling cards.

Step 8: Watch, Comment And Unblock

hermes kanban watch
hermes kanban list
hermes kanban stats

Or run hermes dashboard and open the Kanban tab, a bundled dashboard plugin with drag-drop columns, per-profile lanes inside Running and live WebSocket updates.

From any connected chat, /kanban comment and /kanban unblock go through even while an agent is mid-turn, because the command bypasses the running-agent guard.

A comment lands on the card's thread, and the next run of that card reads it through kanban_show().

For the newsletter, the useful habit is a two-minute check each morning: anything in blocked gets a comment with the decision it is waiting on, and anything in review gets read before it goes out.

If a card blocks twice for the same reason, the docs say repeated same-kind blocks escalate to triage, which is your cue to rewrite the card rather than nudge it again.

Config Keys I'd Check On Day One

KeyDefaultWhy it matters for the planner
kanban.dispatch_in_gatewaytrueKeeps the dispatcher in the gateway; set false on every extra gateway.
kanban.dispatch_interval_seconds60How often ready cards get picked up.
kanban.auto_decomposetrueAuto-splits cards dropped in Triage with the built-in decomposer.
kanban.auto_decompose_per_tick3Caps decompositions per tick so a bulk load doesn't spike usage.
kanban.orchestrator_profileemptyWhich profile owns the root card after a fan-out.
kanban.default_assigneeemptyWhere a child lands if the model picks an unknown profile.
kanban.auto_subscribe_on_createtrueWakes the creating agent when a card it made finishes or blocks.
kanban.failure_limit2Consecutive spawn failures before a card is auto-blocked.

Every default above comes from the Kanban reference; change only the ones you have a reason to change.

Auto Mode: Let Triage Do The Splitting

Sometimes you don't want to brief the planner at all.

Drop a one-line idea such as "plan Q1 client webinar" into the Triage column, and with auto_decompose on, the dispatcher runs the built-in decomposer on its next tick.

It reads your profiles and their descriptions, builds a small graph of child cards and keeps the original card as the parent of every leaf.

When the children finish, the root goes back to ready so its owner can judge whether the work is complete.

The Orchestration: Auto/Manual pill at the top of the board flips this; Manual only stops the built-in Triage decomposer, not your planner's own kanban_create calls.

📞 Planning a multi-agent setup for your own firm? Book a free SEO strategy session and my team will look at where a planner and workers would actually fit your workload. → Book the free session

Running More Than One Gateway

Hermes can run one gateway per profile at the same time.

The multi-gateway guide says only one of them should own the dispatcher, so two gateways never race to spawn the same card.

On every other profile gateway, add this to its config:

kanban:
  dispatch_in_gateway: false

The environment variable HERMES_KANBAN_DISPATCH_IN_GATEWAY=false does the same job.

Those non-dispatch gateways still deliver notifications for their own profiles' chats.

Gotchas That Stall A Board

  • Support card linked under its own blocker. If a blocked worker creates a helper card and links it as a child of the blocked card, neither runs. Put the parent id in the helper's body instead, and free an existing deadlock with hermes kanban unlink <parent> <child>.
  • Workers that end without a verdict. Each run has to finish with kanban_complete, kanban_request_review or kanban_block; anything else is reaped as crashed, gave_up or timed_out.
  • Shared choices left to workers. If two parallel cards both need a file format or naming scheme, the planner decides once and writes it into both bodies.
  • Cross-board links. Each extra board gets its own database under ~/.hermes/kanban/boards/<slug>/, and links between boards are refused.
  • Two homes on one shared database. The default profile name collides by construction; use unique profile names or set kanban.dispatch_profiles per home.

When To Use delegate_task Instead

The reference draws the line clearly: delegate_task is a function call, and Kanban is a work queue.

Delegation blocks the parent, uses an anonymous child and can't resume after a failure.

Use it when the agent needs a quick reasoning answer back inside its own turn.

Use the planner and board when work crosses profiles, must survive restarts, or might need you to step in.

The two also mix: a worker can call delegate_task inside a card, and those children count against delegation.max_concurrent_children and delegation.max_spawn_depth.

And keep /goal separate in your head; the goals page says it loops inside one chat and never creates, claims or moves cards.

For a narrative walk-through with screenshots, the Kanban tutorial runs four stories end to end, and the source lives in the Hermes Agent repository.

Members of the AI Profit Boardroom can bring their board to a live coaching call and get the setup checked step by step.

Also On Our Network

FAQ: Hermes Agent Kanban Orchestrator

What tools does the Hermes Kanban orchestrator use?

kanban_create to fan out cards, kanban_link to add dependencies, kanban_list to read the board, kanban_unblock to restore blocked cards, and kanban_comment on other cards.

Why are my Kanban cards stuck in ready?

Either no gateway is running, so the dispatcher is not ticking, or the assignee name does not match a real profile and the card logged a skipped_nonspawnable event.

Which tools should the planner profile have?

The docs pair it with kanban, gateway and memory only, and the worker lanes page leaves out terminal, file, code and web so it cannot do implementation work.

How often does the dispatcher pick up cards?

Every 60 seconds by default, set by kanban.dispatch_interval_seconds. The dashboard also has a Nudge dispatcher button so you do not have to wait.

Can I run several gateways with one board?

Yes. Leave dispatch_in_gateway true on one gateway and set it to false on the others. The others still deliver notifications for their own profiles.

📺 Video notes + links to the tools 👉

🎥 Learn how I make these videos 👉

🆓 Get a FREE AI Course + Community + 1,000 AI Agents 👉

About Julian

I'm Julian Goldie, an SEO entrepreneur, author and founder of the AI Profit Boardroom, which has 3,400+ members.

I help business owners scale with AI agents, automation and SEO.

  • I've built a 7-figure SEO and link building agency, Goldie Agency, with a team of 70+ SEO experts.
  • I've grown a YouTube channel to 400,000+ subscribers.
  • I wrote "Link Building Mastery".
  • My Udemy courses have taught over 29,000 students.

→ Get my best AI training inside the AI Profit Boardroom

Create the worker roster, lock the planner to kanban, gateway and memory, start the gateway, and if you want my team to sanity-check your plan, book a free strategy session and bring your hermes agent kanban orchestrator.

§ Field reports

Real wins from inside the AI Profit Boardroom

See all 3,400+ members →
AIPB member win screenshot AIPB member win screenshot AIPB member win screenshot AIPB member win screenshot AIPB member win screenshot AIPB member win screenshot AIPB member win screenshot AIPB member win screenshot AIPB member win screenshot AIPB member win screenshot AIPB member win screenshot AIPB member win screenshot

Scroll sideways →

⚑ Next step

Ready to run AI agents that actually do the work?

Join 3,400+ operators inside the AI Profit Boardroom. Get 1,000+ plug-and-play AI agent workflows, 5 coaching calls a week, and a community that holds you accountable.

Join The AI Agent Community →

7-day no-questions refund · Cancel anytime

Found this useful? Get more of this site in your Google.

One tap adds us as a preferred source — Google then shows you our posts more often in Search, Top Stories, and AI Overviews.

← Back to all posts