Hermes Agent Kanban: Our Daily AI Video Team of Four Agents
Summary
Inside our Hermes Agent Kanban setup: one scheduled daily card fans out to research, script, video-generation and publishing agents that ship a finished YouTube video end to end. Exact board design and CLI included.
Producing one good video a day is not a creativity problem, it is a coordination problem — research, scripting, rendering and publishing are four different jobs with four different failure modes, and a single AI agent doing all of them in one long session degrades badly. We solved it with the Kanban board in Hermes Agent: one scheduled card per day fans out to a research agent, a script agent, a video-generation agent and a publishing agent, and a finished video lands on the channel without anyone touching a terminal. This post shows the exact board design and CLI configuration. If you want the same pipeline built for your team, book a free 30-minute consultation.
Here is a video produced end to end by this agent team — no human edited a frame of it:
Why single-agent video pipelines fall over
Our first daily-video automation ran as one monolithic agent loop, and we documented it in an earlier Hermes Agent video workflow post. It worked, but it had a structural weakness: everything lived in one context window. By the time the agent reached the upload step it was reasoning over hours of accumulated research notes, script drafts and render logs, and quality suffered exactly where it matters most — the final cut. When a render failed at 2am, the whole chain restarted from ideation.
Nous Research shipped the fix in Hermes Agent v0.12: a Kanban multi-agent board where named agent profiles claim tasks, work in parallel, and hand off through a durable comment thread instead of a shared context window. Every task is a row in a SQLite database (~/.hermes/kanban.db); every handoff is a row any agent can read. A dispatcher scans the board on a configurable interval, promotes tasks whose dependencies are done, and spawns the assigned profile with its own tools and memory. Fittingly, Nous open-sourced the whole system and used the Kanban itself to plan and produce its own launch video.
How the Kanban turns one daily card into a four-agent pipeline
The board model
Tasks flow through a fixed lifecycle: triage → todo → ready → running → done, with blocked available at any point for human escalation. The two primitives that make a production pipeline possible are:
- Profiles — named worker identities, each with its own system prompt, toolset and persistent memory. Our board has four:
researcher,scriptwriter,videogenandpublisher. - Parent dependencies — a card only becomes dispatchable when all of its parents reach
done. Chain them and you get a pipeline; give one card several parents and you get a fan-in.
The daily trigger
A cron job creates the day's parent card. The --idempotency-key flag matters — if cron ever double-fires or a retry runs, the board deduplicates instead of producing two videos:
hermes kanban create "daily video: research today's topic" \
--assignee researcher \
--scheduled-at "2026-07-16T22:00:00Z" \
--idempotency-key "video-2026-07-16" \
--max-runtime 45mThe dispatcher skips future-dated cards until scheduled-at passes, so the run starts overnight Singapore time and the finished video is live before the workday begins.
The dependency chain
The remaining three cards are created up front, each parented on the previous stage:
hermes kanban create "write 90-second script from research brief" \
--assignee scriptwriter --parent t_research
hermes kanban create "render video from approved script" \
--assignee videogen --parent t_script --max-retries 2
hermes kanban create "upload, title, description, thumbnail, publish" \
--assignee publisher --parent t_renderEach agent finishes by calling kanban_complete with machine-readable metadata — the researcher hands off source URLs and a fact list, the scriptwriter hands off the final script and scene notes, the video agent hands off the rendered file path and duration. The next agent reads the full comment thread on spawn, so context transfers without any shared session. If the render fails twice, the card flips to blocked and we get pinged to unblock from the dashboard — the only human touchpoint in the whole system.
What each agent actually does
| Agent profile | Job | Handoff metadata |
|---|---|---|
researcher | Searches the day's AI news, picks one topic, verifies 3–5 sources | Topic, angle, fact list, source URLs |
scriptwriter | Turns the brief into a timed 90-second script with scene directions | Script text, scene list, voiceover notes |
videogen | Generates visuals and voiceover, assembles and renders the cut | File path, duration, render log |
publisher | Writes title, description and tags, uploads, schedules the premiere | Video URL, publish time |
Concurrency is capped in config (max_in_progress_per_profile), heartbeats reclaim crashed workers automatically, and a failure_limit circuit-breaker stops a bad day from burning tokens in a loop. These are the operational guardrails that separate a demo from something you can leave running for months — the same discipline we apply in our AI agent deployment engagements.
Single agent loop vs Kanban team — what changed for us
| Dimension | One agent, one loop | Kanban team of four |
|---|---|---|
| Context quality | Degrades over a long session | Fresh context per stage, thread handoffs |
| Failure recovery | Restart the whole run | Retry one card; upstream work is preserved |
| Observability | One long log to scroll | Live board — see which stage is stuck at a glance |
| Specialisation | One generic prompt does four jobs | Each profile has a tuned prompt and toolset |
| Human touchpoints | Babysit the terminal | Unblock from the dashboard only when escalated |
The practical difference shows up in the archive: our blocked-card rate is a handful per month, and every one of them was a genuine editorial judgement call rather than a mechanical failure. If you are still choosing an agent stack, our OpenClaw vs Hermes vs Paperclip comparison covers where Hermes fits against the alternatives. Want to see the board running live against your own use case? Book a 30-minute walkthrough and we will run a card through all four agents while you watch.
Our approach: pipelines you can hand to an operations team
We build and deploy this pattern commercially as part of our custom AI solutions practice — the video team above is our own dogfooded instance, not a lab demo. A typical engagement scopes the stages, writes and red-teams each profile's system prompt, wires the schedulers and circuit-breakers, and hands over a dashboard your operations staff can supervise without reading a line of code.
For teams that want to build this capability in-house, the skills are trainable. The WSQ Build a Human-AI Workforce with Autonomous AI Agents course covers exactly this operating model — designing agent teams, supervision gates and handoffs — with hands-on labs on exactly the kind of agent team this pipeline runs. The course is WSQ-funded for Singapore companies.
FAQ
Do the agents ever publish something wrong?
The publisher profile enforces a checklist before upload — script facts must trace to the researcher's source list, and any mismatch flips the card to blocked instead of publishing. In months of daily runs, escalations have been rare and none reached the channel. You choose where the human gate sits: fully autonomous, or a mandatory review card between render and publish.
What does a daily video actually cost to run?
The dominant cost is video-generation inference; orchestration overhead on the Kanban is negligible because the board is a local SQLite file and the dispatcher is a lightweight poller. A 90-second daily video typically runs at a small fraction of what a freelance editor would charge for one video a week.
Does this need Hermes specifically?
The pattern — scheduled parent card, dependency chain, metadata handoffs, circuit-breakers — is portable. Hermes Agent is currently the cleanest open-source implementation of it, but we have shipped the same architecture on other stacks. See our human–AI collaboration deep dive for how the supervision model compares across agents.
Can this pipeline produce training or marketing videos instead of news explainers?
Yes — the stage structure is content-agnostic. Swap the researcher's brief for your course syllabus or product changelog and the rest of the chain is unchanged. That is usually the first customisation we scope in a consultation.
What to do next
- Read the official Hermes Agent Kanban documentation and run the tutorial board locally — it takes an evening.
- Learn to run agent teams like this one on the funded WSQ Build a Human-AI Workforce with Autonomous AI Agents course, or browse the wider AI courses at Tertiary Courses Singapore.
- Build — request a scoped quote for a multi-agent content pipeline on your own channel, LMS or product.