Hermes Agent Kanban: How a Team of AI Agents Produced Our YouTube Video

One human, four AI agents, one finished YouTube video

Most teams use AI as a single assistant: one chat window, one task at a time. We wanted to test the next step — a team of autonomous AI agents that divides up a real production job the way a human team would. The job: research, script, produce and publish a YouTube video for our channel. The coordination layer: the Hermes Agent Kanban board.

The result is the video below — researched, scripted, produced and published by the agent team, with a human supervising the board rather than doing the work.

What is the Hermes Agent Kanban board?

Hermes Agent (from Nous Research) added a feature that changes how multi-agent work is organised: a durable, shared Kanban board that multiple named agents work from. Instead of fragile in-process subagent swarms that vanish when a session dies, tasks live as cards in a persistent board with a title, an assignee (the agent profile that should do the work) and a status.

• A dispatcher loop ticks in the background, claims cards that are ready, and spawns the assigned agent in its own clean workspace
• Dependencies are first-class — a card can wait for upstream cards to finish, so downstream agents start automatically the moment their inputs are ready
• Crash recovery is built in — if an agent dies mid-task, the board reclaims the card and respawns the worker
• The full run history is preserved, so you can audit what every agent did and why

In short: it turns multi-agent coordination from prompt spaghetti into a project board your agents actually work off — the same mental model every operations team already knows.

Our four-agent video production team

We defined four named agent profiles and put four dependent cards on the board:

• Research Agent — investigates the topic (in this run: Microsoft retiring native SMS login and how passkeys stop phishing), gathers sources, and produces a fact sheet with citations
• Script Agent — waits on the research card, then turns the fact sheet into a tight narration script with a hook, structure and call to action
• Video Agent — waits on the script card, then generates the voiceover and visuals and renders the final video file
• Publishing Agent — waits on the video card, then writes the SEO title, description and tags, and publishes to the YouTube channel

Because dependencies are handled by the board, nobody 'hands off' anything manually. The Script Agent wakes up the moment research is done; the Publishing Agent fires the moment the render lands. The human's job shrinks to what it should be: set the goal, review the cards, approve the output.

Why this matters for your business

This is the shape of the human-AI workforce: not one person prompting one chatbot, but one person supervising a board where specialist agents do the research, drafting, production and publishing. The same pattern applies far beyond video — think lead research handed to an outreach-drafting agent, or report data handed to an analysis agent and then a formatting agent. The skill your team needs is no longer prompting — it is designing, orchestrating and governing agent teams.

Learn to build your own agent workforce — WSQ funded

Our WSQ Build a Human-AI Workforce with Autonomous AI Agents course teaches exactly this: designing autonomous agent teams, orchestrating multi-agent workflows, and governing them safely in a real organisation. As a WSQ-accredited course it attracts up to 70% funding for eligible Singaporeans and PRs, individuals can offset fees with SkillsFuture Credit, and companies can stack SkillsFuture Enterprise Credit (SFEC) plus absentee payroll support. View class dates and register here.

FAQ

Do I need to be a developer to run a multi-agent team?

No — the Kanban model is deliberately non-technical: cards, assignees, statuses and dependencies. The course starts from fundamentals and builds up to orchestrating full agent teams.

Is a multi-agent board better than one powerful agent?

For multi-step jobs, yes — specialist agents with clear handoffs are more reliable and auditable than one agent juggling everything, and the durable board survives crashes that kill in-process swarms.

Is the course funded?

Yes — WSQ funding up to 70%, SkillsFuture Credit for individuals, and SFEC for company-sponsored staff.

What to do next

1. Pick one multi-step job your team repeats — a report, a campaign, a video
2. Sketch it as four or five cards with clear handoffs, exactly like our video pipeline
3. Reserve seats in the WSQ Build a Human-AI Workforce with Autonomous AI Agents class and leave able to run that board with real agents