Video: "New Hermes Agent Kanban Update is INSANE!" by Julian Goldie on YouTube.
What the Kanban board actually is
The Kanban update adds a visual task board — structured like a Trello board — directly inside Hermes Agent. Tasks sit in columns (to do, in progress, done), and a dispatcher process checks the board continuously. When a task is ready, the dispatcher launches the appropriate agent profile, assigns the task, and updates the card as work progresses.
This is not just a visual layer on top of the existing single-agent setup. It is a fundamentally different model of work. Previously, one Hermes agent handled tasks in sequence. Now multiple agents can be live simultaneously, each working on a different card, all from the same shared board.
How the dispatcher coordinates agents
The dispatcher is the key piece most coverage glosses over. It is a lightweight background process that polls the board, checks which tasks are in a ready state, and matches each task to a named agent profile. Each profile has specific tools, memory, and instructions — so the research agent and the writing agent do not collide.
When an agent finishes, the card moves. When an agent gets stuck, it can flag the task for human review or pass it to a different profile. In practice this means a failed subtask does not stall the whole job the way it did before.
What changes for longer jobs
The practical improvement shows on anything that used to require chaining prompts manually. A content workflow that previously needed you to run keyword research, then brief writing, then HTML generation as three separate agent runs can now sit on one board and run end-to-end. You drop the job in, check back later, and the cards have moved.
Worth knowing: parallel agents share the same underlying model API, so there is a cost consideration. If you are running five agents simultaneously against GPT or Claude, you will burn tokens five times as fast. For many tasks that trade-off is worth it; for simple jobs it is not.
The limits to be realistic about
Multi-agent Kanban is not autonomous project management. The dispatcher still needs sensible task definitions — vague cards produce vague output. Inter-agent handoffs require careful agent profile design: if the research agent produces output the writing agent does not know how to parse, the workflow stalls at the seam.
There is also no built-in conflict resolution when two agents both need the same resource. In practice you design around this by scoping tasks carefully, but it does need thinking through before you run anything in production.
Where this connects to NordSys
Properly designed multi-agent workflows are not something you bolt on as an afterthought. The value is in the design: sensible task decomposition, well-scoped agent profiles, and a board structure that matches how your business actually works. We help clients build exactly this — from the initial agent setup through to production workflows that keep running without constant babysitting. Our AI Agents service covers the full setup and ongoing support.
See our AI Agents service →