My hermes agent kanban orchestrator verdict: it is the cleanest way Hermes offers to run several named agents on one plan, and it scores 8.2 out of 10 on my card, held back mainly by a setup that punishes small naming mistakes.
The orchestrator is a Hermes profile with the Kanban toolset that decomposes a goal into board tasks, links them, assigns each to an existing profile and steps back.
Below is what was reviewed, the scorecard, the strengths, the limits, who should use it, who shouldn't, and the final call.
Want my team to look at whether a multi-agent board fits your business before you build one? Book a free SEO strategy session.
What Was Reviewed
This verdict rates the feature from the official documentation only: the Kanban reference, the worker lanes contract, the multi-gateway deployment page, the Kanban tutorial and the goals page.
Ratings are editorial judgement based on what those pages say the feature does and how it fails.
Where the docs give a hard default, such as a tick interval or a failure limit, the number is quoted as written.
No benchmark was run for this review, and no model is recommended, because the docs themselves use placeholder model names.
The tutorial matters for the rating because it shows the board under stress: a solo developer shipping a feature, fleet work across many subjects, a role pipeline with a review retry, and a circuit breaker with crash recovery.
Those four stories are where the failure handling stops being theory.
The Scorecard
| Criterion | Rating | Reason from the docs |
|---|---|---|
| Role separation | 9/10 | The orchestrator lane strips implementation tools, so the planner cannot quietly do the work. |
| Durability | 9/10 | Tasks, handoffs and runs are SQLite rows that survive restarts; crashed workers are reclaimed. |
| Human control | 9/10 | Comment, block, unblock and review at any point, including from chat apps mid-turn. |
| Dependency handling | 8/10 | Children wait in todo until every parent is done; one linking mistake can still deadlock two cards. |
| Setup effort | 7/10 | A few commands, but the toolset is off by default and unknown assignee names fail silently. |
| Auto orchestration | 8/10 | A built-in decomposer splits Triage cards by profile description, capped at 3 per tick. |
| Dashboard | 9/10 | Bundled plugin with drag-drop, per-profile lanes, live updates and bulk actions. |
| Scale-out | 6/10 | Single-host by design; shared boards across homes work only with careful profile naming. |
| Documentation | 9/10 | A long reference plus a four-story tutorial; one small wording conflict on how guidance ships. |
| Overall | 8.2/10 | Average of the nine criteria above. |
Strength: The Planner Is Built To Stay A Planner
Most multi-agent setups fail the same way: the coordinator gets impatient and does the job itself.
Hermes attacks this from two sides.
First, the reference says anti-temptation rules, a profile-discovery step and a decomposition playbook are injected into the system prompt automatically.
Second, the worker lanes contract defines an orchestrator profile whose toolset includes kanban but excludes terminal, file, code and web.
The docs' own line sums up the design goal: "A well-behaved orchestrator does not do the work itself."
That earns the top-tier score for role separation.
Strength: Durable Work Instead Of Fragile Subagents
The reference compares Kanban with delegate_task directly, and the gap is large.
Delegation is a blocking call to an anonymous child with no resume; Kanban is a queue where named profiles pick up cards, can be blocked and retried, and leave a permanent audit trail.
Every run must end with kanban_complete, kanban_request_review or kanban_block, and anything else is reaped as crashed, gave_up or timed_out.
After 2 consecutive spawn failures, the default kanban.failure_limit, the card is auto-blocked with the last error as its reason, so a broken worker doesn't loop.
Open-ended cards get a further option: --goal runs that card's worker in a judge loop, where an auxiliary judge checks each turn against the card's title and body as acceptance criteria.
The default budget is 20 turns, and running out blocks the card for human review instead of closing it quietly.
If the judge rules the goal unachievable as written, the card is blocked straight away, so an impossible card is never marked done.
That is a sensible guard for planners that write vague cards, and it is one reason durability scores a 9.
๐ฅ Want to see a working planner-and-workers board before you rate it yourself? Inside the AI Profit Boardroom you get step-by-step video trainings, the installable Agent OS with Hermes, four live coaching calls a week and 3,600+ members comparing agent setups. โ Join the AI Profit Boardroom
Strength: You Can Steer From Anywhere
/kanban takes the same verbs and flags as the CLI and works inside hermes chat and on gateway platforms including Telegram, Discord, Slack, WhatsApp, Signal, Matrix, Mattermost, email and SMS.
It bypasses the running-agent guard, so a comment or an unblock lands immediately even while an agent is mid-turn.
Creating a card from a chat subscribes that chat to the card's terminal events, so the result comes back to the thread that asked for it.
The one catch is message length: output over about 3,800 characters is truncated in chat apps.
If you already run Hermes in a team chat, the Hermes Agent Slack verdict rates that side of the setup.
Strength: The Dashboard Is Part Of The Install
Run hermes dashboard and a Kanban tab appears, shipped as a bundled plugin rather than a separate service.
It shows one column per status, per-profile lanes inside Running, a create-task dialog with skills, goal mode and parent, multi-select bulk actions and live updates over WebSocket.
Moves into done, archived or blocked ask for confirmation, and a Nudge dispatcher button saves waiting for the 60-second tick.
Limit: Setup Has Silent Failure Points
A regular chat profile has no kanban_* tools until you run hermes -p <profile> tools enable kanban and start a new chat.
Worse, a typo in an assignee gives no error: per the reference, a card whose assignee doesn't resolve stays in ready and records a skipped_nonspawnable event.
The planner's injected guidance includes a profile-discovery step for this reason, but a human typing --assignee by hand gets no such help.
That is why setup effort scores 7, not 9.
Limit: One Linking Mistake Can Deadlock Two Cards
The reference carries a caution box on this.
If a blocked worker creates a support card and links it as a child of the card it is meant to unblock, the support card waits behind its own blocker and neither runs.
The board does flag it with gated: true and a dependency_wait event, and hermes kanban unlink releases it.
Visible, recoverable, but easy to trigger.
Limit: Built For One Machine
The docs describe Kanban as single-host by design.
Several gateways can run at once, but the multi-gateway page says only one should own the dispatcher, with kanban.dispatch_in_gateway: false on the rest.
Mounting one kanban.db in several Hermes homes is possible, yet every home has a default profile that collides, so you need unique names or kanban.dispatch_profiles per home.
Boards themselves are hard walls: each has its own database, workers see only their own board, and cross-board links are refused.
That keeps things simple, and it is why scale-out gets the lowest score on the card.
One more practical catch sits in workspaces.
The default scratch workspace is a fresh temp folder that is deleted when the card completes, keeping only files the worker declares as artifacts.
For anything you want to keep whole, the reference points to dir: with an absolute path or a worktree for code, both preserved on completion.
It is documented clearly, but first-time users will lose a draft or two before it sinks in.
Auto Mode, Rated On Its Own
With kanban.auto_decompose on by default, the dispatcher runs a built-in decomposer on Triage cards, up to 3 per tick.
It reads profile descriptions to route children and never leaves a child unassigned: unknown picks fall back to kanban.default_assignee, then the root's assignee, then the active default profile.
The original card stays as the parent, so its owner wakes to judge completion once the children finish.
Quality depends on your descriptions, which is fair, and the Auto/Manual pill gives you a quick off switch.
Manual mode has a clear boundary in the docs: it only stops the built-in Triage decomposer, and your planner can still create cards whenever it likes.
There is also hermes kanban specify for rewriting one rough card into a proper spec without fanning it out.
Who Should Use It
- Teams with repeat work that naturally splits into roles, such as research, drafting and review.
- Anyone who needs work to survive restarts, or wants a record of every attempt and handoff.
- Owners who want to approve or redirect work from a phone without opening a terminal.
- Builders who already run several Hermes profiles and want one board over all of them.
- Anyone who wants a review step baked in, since workers can hand a card to review and reviewers can request changes without it counting as a block.
Who Shouldn't
- Anyone whose AI work is one task at a time;
/goalkeeps a single chat going until a judge says done, and it never touches the board. - People who need a quick answer inside one turn;
delegate_taskis lighter for that. - Setups that need a distributed queue across many machines, which the docs don't claim to offer.
๐ Weighing a board like this for your own workload? Book a free SEO strategy session and my team will look at which of your jobs suit a planner and workers, and which don't. โ Book the free session
Final Verdict
8.2 out of 10.
The orchestrator gets the hard part right: it is designed and fenced so planning and doing stay separate, and every handoff lands on a durable board people can read.
The weak spots are practical rather than conceptual: a toolset that's off by default, silent failures on wrong names, and a one-host ceiling.
Fix those with a profile check before every first run and the score feels conservative.
The source and release notes live in the Hermes Agent repository if you want to follow changes.
For guided setups and live reviews of real boards, the AI Profit Boardroom runs Hermes sessions on its weekly calls.
Also On Our Network
- ๐ The eight-step setup for a Hermes Kanban planner profile and its workers
- ๐ Whether a small business should run a Hermes Kanban planner and worker board
- ๐ Five copy-paste prompts to run a Hermes Kanban planner inside your Agent OS
- ๐ Which AI models to pair with a Hermes Kanban planner and its workers, by real GoldieBench scores
FAQ: Hermes Agent Kanban Orchestrator Verdict
What score does the Hermes Kanban orchestrator get?
8.2 out of 10, the average of nine criteria rated from the official docs. Scale-out is its weakest area at 6 out of 10.
Is the orchestrator better than delegate_task?
They do different jobs. delegate_task is a blocking call for a quick answer; the Kanban orchestrator suits work that crosses profiles, must survive restarts or needs human input.
What is the biggest setup trap?
Assignee names that do not match a real profile. The card waits in ready with a skipped_nonspawnable event instead of running.
Does it work across several machines?
The docs describe Kanban as single-host by design. Sharing one board across Hermes homes is possible but needs unique profile names or dispatch_profiles per home.
๐บ Video notes + links to the tools ๐
๐ฅ Learn how I make these videos ๐
๐ Get a FREE AI Course + Community + 1,000 AI Agents ๐
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.
โ Get my best AI training inside the AI Profit Boardroom
Enable the toolset on one planner, check every assignee name, start the gateway and judge the board on a week of real work, and if you want a second view, book a free strategy session and bring your hermes agent kanban orchestrator.











