This article uses the example of generating a modern, technology-themed Mr. Bean TV episode script based on current global events to demonstrate the Kanban-based multi-agent collaboration design pattern. The article uses Hermes Agent, but the same approach can also be applied to other agent frameworks or even your own custom agents.
Introduction
For complex AI tasks, the question is no longer whether multiple agents should collaborate, but how to make multi-agent collaboration work effectively.
The short-lived “spawn agent” model has several obvious problems:
- Limited controllability
- Poor observability
- Ephemeral state and context loss
Kanban, a project collaboration method that has proven effective across many industries, can also be applied to multi-agent collaboration.
- Kanban makes task states observable
- Kanban works naturally with Human-in-the-Loop workflows
For example:
- Manually resolving blocking issues
- Triggering retries
- Providing small amounts of critical context
- Controlling and supervising task execution
Kanban
Toyota’s Kanban method originated from post-war manufacturing practices, especially the production management system Toyota developed during the 1940s and 1950s.
The word “Kanban (看板 / かんばん)” originally meant:
- Signboard
- Notice board
- Storefront sign
Later, the meaning evolved into “a card or marker used to transfer information.”
Why does “kanban” look similar to Chinese pinyin? Because it is originally a Japanese word written with Chinese characters (“看板”), pronounced “kanban” (かんばん) in Japanese. Many Japanese words historically originated from Chinese characters.
Kanban Application Example
Let’s start with a practical example before discussing the underlying principles. This makes the concept easier to understand intuitively.
Planner Goal
In Hermes, enter the following prompt:
Use the Kanban method to generate agent tasks and dependency relationships:
- Task 1: Historical research
- Research the history and style of the Mr. Bean TV series and save it as `mr-bean-history.md`
- Task 2: Main actor status research
- Research the current status of the lead actor and save it as `mr-bean-status.md`
- Task 3: Major world events in 2026
- Research major world events in 2026 and save the result as `mr-bean-2026-events.md`
- Final Task:
- Design a 2026 version of a Mr. Bean TV episode story script and save it as `mr-bean-2026-story.md`
- This task depends on Task 1, Task 2, and Task 3
Assign all tasks to the `default` profile
Output:
✅ Kanban tasks created successfully
Successfully created 4 tasks. Tasks 1, 2, and 3 are independent and can run in parallel. Task 4 (the final task) depends on the first three tasks.
📋 Task List
ID Status Dependencies Description
────────── ─────── ──────────── ────────────────────────────
t_129bb524 ✅ Ready None Task 1: Mr. Bean history and style research
t_339dcadc ✅ Ready None Task 2: Lead actor status research
t_f55ad7f6 ✅ Ready None Task 3: 2026 world events research
t_b1a7f48d 🔒 Todo Task1,2,3 Final task: 2026 Mr. Bean episode script design
🔄 Task Dependency Graph
Task1 (ready) ─┐
Task2 (ready) ─┼→ Final Task (todo)
Task3 (ready) ─┘
• Tasks 1, 2, and 3 can run in parallel
• The final task will start automatically after all three prerequisite tasks are completed
Viewing Tasks
You can view the Kanban board in a browser:
hermes dashboard
Final Generated Script
mr-bean-2026-story.md:
# Mr. Bean in 2026 — TV Episode Script
> **Title:** Mr. Bean: Smart Beans
> **Duration:** Approximately 30 minutes (single episode)
> **Style Direction:** Classic physical comedy combined with 2026-era technology themes. Considering Rowan Atkinson is already 71 years old, high-intensity physical actions (chases, wrestling, dangerous stunts) are significantly reduced. The humor focuses more on "intelligent awkwardness" — carefully designed pauses, subtle body movement mismatches, facial micro-expressions, and interactions with modern smart devices.
---
## Story Outline
Mr. Bean moves into a newly upgraded smart apartment (12 Arbour Road, renovated by his landlady). He must survive an entire day surrounded by AI assistants, automatic doors, voice-controlled lighting, and robotic housekeepers — while also trying to attend a 2026 FIFA World Cup qualifying match (England vs. Brazil).
The entire episode revolves around Bean’s disastrous interactions with “overly intelligent” modern technology, structured around three parallel storylines:
- **Plot A (At Home):** Bean’s war against smart home technology
- **Plot B (On the Road):** Bean’s Mini encounters self-driving cars and fuel price chaos
- **Plot C (Outside the Stadium):** Bean’s misadventures at the FIFA World Cup experience zone
**Design Principle:** Minimal dialogue throughout the episode. Mr. Bean barely speaks, aside from his iconic mumbling sounds. The humor comes from his childlike curiosity, naive misunderstandings of technology, and overconfident problem-solving attempts.
---
## Opening Sequence
**Classic Animated Opening (Updated Version):**
A glowing bean pod falls from space against the backdrop of NASA Artemis II’s lunar mission trajectory (referencing the crewed moon mission planned for April 2026). The pod crashes through clouds and lands on the solar panels of a smart apartment building. Bean pops out of the pod holding Teddy, who is now wearing a VR headset, and lands on the balcony of Flat 2.
The opening music remains Howard Goodall’s iconic theme song, but this time the choir is generated by AI, referencing the modern “AI choir” trend. The lyrics remain unchanged: *Ecce homo qui est faba*.
---
## Scene One: Bean’s Smart Apartment (Approx. 8 Minutes)
### Setting
Bean wakes up in the morning. His apartment has been upgraded by his landlady into a “fully smart home.” Touchscreens cover the walls, gesture-controlled lights hang from the ceiling, and a small robot housekeeper — resembling a robotic vacuum cleaner — waits in the kitchen.
### Story Development
#### 1.1 The Wake-Up Battle (2 Minutes)
Bean’s beloved vintage mechanical alarm clock has been replaced by a rounded AI voice assistant called “BEE-BOP” (Bean's Environment & Entertainment Brain-Operated Platform — a random name invented by the landlady, which Bean does not understand).
...
Example Summary
As you can see, if I simply used a single prompt such as:
“Generate a modern, technology-themed Mr. Bean episode script based on current global trends”
the result from the agent would likely be very different.
More importantly, it probably would not perform parallel research tasks automatically.
Throughout the process, task states and logs can be monitored in the Kanban dashboard or through:
hermes kanban watch
During my own execution, some tasks even failed multiple times because of network issues and eventually entered the block state. I had to manually resolve the problem and then change the task state from block back to ready.
This is exactly where Human-in-the-Loop workflows become valuable.
Kanban Principles
First, let’s look at the data model:
Task states:
Architecture and workflow:
Kanban Task Execution Protocol
Each worker’s system prompt includes the following Kanban-related instructions:
from agent.prompt_builder import KANBAN_GUIDANCE
# Kanban task execution protocol
You have been assigned ONE task from the shared board at `~/.hermes/kanban.db`. Your task id is in `$HERMES_KANBAN_TASK`; your workspace is `$HERMES_KANBAN_WORKSPACE`. The `kanban_*` tools in your schema are your primary coordination surface — they write directly to the shared SQLite DB and work regardless of terminal backend (local/docker/modal/ssh).
## Lifecycle
1. **Orient.** Call `kanban_show()` first (no args — it defaults to your task). The response includes title, body, parent-task handoffs (summary + metadata), any prior attempts on this task if you're a retry, the full comment thread, and a pre-formatted `worker_context` you can treat as ground truth.
2. **Work inside the workspace.** `cd $HERMES_KANBAN_WORKSPACE` before any file operations. The workspace is yours for this run. Don't modify files outside it unless the task explicitly asks.
3. **Heartbeat on long operations.** Call `kanban_heartbeat(note=...)` every few minutes during long subprocesses (training, encoding, crawling). Skip heartbeats for short tasks.
4. **Block on genuine ambiguity.** If you need a human decision you cannot infer (missing credentials, UX choice, paywalled source, peer output you need first), call `kanban_block(reason="...")` and stop. Don't guess. The user will unblock with context and the dispatcher will respawn you.
5. **Complete with structured handoff.** Call `kanban_complete(summary=..., metadata=...)`. `summary` is 1–3 human-readable sentences naming concrete artifacts. `metadata` is machine-readable facts (`{changed_files: [...], tests_run: N, decisions: [...]}`). Downstream workers read both via their own `kanban_show`. Never put secrets / tokens / raw PII in either field — run rows are durable forever.
6. **If follow-up work appears, create it; don't do it.** Use `kanban_create(title=..., assignee=<right-profile>, parents=[your-task-id])` to spawn a child task for the appropriate specialist profile instead of scope-creeping into the next thing.
Conclusion
Kanban-style multi-agent collaboration can be applied to many use cases, including:
- Research content generation
- Software development
- Workflow orchestration
- AI-assisted project management
One of the biggest advantages is that Kanban is already widely understood by humans. The learning curve is relatively low, and many mature management methodologies and operational practices can be directly reused in AI agent systems.
Last but not least, I am currently open to new job opportunities. If you think my background may be a good fit, feel free to reach out:
https://blog.mygraphql.com/zh/
