Kanban evolved from sticky notes on a whiteboard to a durable SQLite-backed multi-agent work queue. This guide teaches you what it is, why it matters, and how to put it to work with four real project blueprints.
Kanban is a workflow visualization system. The word means "signboard" in Japanese — it originated at Toyota factories in the 1940s as a way to signal when more parts needed to be produced. Workers would move a physical card from one bin to another, and that card was the signal: make more of this part.
The core idea is simple but powerful: make work visible. Instead of a to-do list where everything blends together, Kanban splits work into columns that represent stages. You can see at a glance what's being worked on, what's waiting, and what's done.
New tasks that need to be fleshed out before they can be worked on. A one-liner idea waiting to become a real spec.
Tasks with a clear spec, but waiting on a dependency (a parent task) to finish first before they can start.
Tasks that are fully specified and unblocked — ready for someone (or something) to pick them up and start working.
Work that is actively being done right now. In Hermes, this means an agent process has been spawned and is executing.
Work that hit a wall and needs a decision, input, or review before it can continue. Human-in-the-loop checkpoint.
Completed work with a structured handoff — summary, metadata, and audit trail preserved forever in the database.
For 80 years, Kanban was a human-only system. The board was the tool, but humans did every single piece of work. A developer moved a card from "Todo" to "In Progress" — then wrote the code. A reviewer moved it to "Review" — then reviewed it. Every column required a person's hands and brain.
Physical cards on a factory floor. A card moves from bin A to bin B → workers produce more parts. The signal IS the system. No software, no dashboards — just paper and discipline.
Digital boards replace physical walls. Trello, Jira, Asana, GitHub Projects. Columns became lists. Cards became tickets. But humans still did every single task — the board was a passive tracking tool, not an active worker.
Dozens of Kanban tools compete on UI polish, integrations, and collaboration features. But the fundamental model is unchanged: the board tracks work, humans do work. Automation is limited to moving cards on a trigger ("when PR merged → move to Done"). The board never does the work itself.
Tools like Claude Code, Codex, and Hermes Agent introduce autonomous agents that can actually do work, not just track it. But most agent tools are stateless — they run, produce output, and forget. The board and the agent are separate systems. You still manually copy agent output back into your project.
Every traditional Kanban tool shares one fundamental constraint: the board is passive. It tracks work, but it never does work. Every card requires a human to pick it up, do the task, and move it forward. If your team is sleeping, work stops. If someone quits, their cards sit frozen until someone else picks them up.
In the AI agent era, this constraint is no longer necessary. The agent can be the worker — but only if the board and the agent are the same system.
Hermes Agent (by Nous Research) introduced Kanban in v0.18.0 (July 2026). What makes it fundamentally different from every Kanban tool that came before is one architectural decision: the board and the agents are the same system.
When you create a task and assign it to a profile (e.g., "research" or "dev"), the dispatcher — a loop running inside the Hermes gateway — automatically spawns a full Hermes Agent process for that profile. The agent reads the task, does the work using its tools (terminal, file, web search, code execution, API calls), and writes the result back to the board. No human intervention required unless the task blocks on a decision.
| Dimension | Traditional Kanban (Trello, Jira, Linear) | Hermes Kanban |
|---|---|---|
| Who does the work | Humans only | AI agent profiles — each a full OS process with tools |
| When work happens | Business hours, when someone picks up the card | 24/7 — dispatcher ticks every 60 seconds |
| Survives crashes | No — if Jira crashes, you lose context | Yes — SQLite database, block/unblock/re-run, crash recovery |
| Memory | None — each card is isolated | Each profile has persistent memory across tasks |
| Human in the loop | Manual — you check the board | Built-in — agent blocks, you get notified, you comment, agent resumes |
| Audit trail | Limited — card activity log | Permanent — SQLite rows with full run history, summaries, metadata |
| Task dependencies | Manual — you remember which task depends on which | Automatic — parent→child links, auto-promotion when parents complete |
| Multi-agent coordination | Impossible — it's a single-user tracking tool | Native — multiple named profiles collaborate on the same board |
| vs. delegate_task | N/A | delegate_task is a function call (RPC); Kanban is a durable work queue that survives restarts |
hermes kanban <verb> — 38+ subcommands from your terminal. Create, list, show, assign, dispatch, complete, archive.
/kanban <action> — from any chat session or gateway platform (Telegram, Discord, Slack, WhatsApp, Signal, email, SMS).
Visual board at port 9119 — six columns, drag-and-drop cards, run history, recovery actions, live updates via WebSocket.
--tenant for clean isolation on the same board.--goal so the worker keeps going until acceptance criteria are met.kanban_complete(summary=..., metadata=...) is the primary channel between stages. Downstream workers read parent summaries automatically.--skills to force-load specific skills into the spawned worker (e.g., translation, code review, security audit).--workspace worktree: so each task gets an isolated git branch.This is a live, interactive simulation of the Hermes Kanban board. Drag cards between columns to see how tasks flow through the pipeline. Each card shows the title, assigned profile, and priority — just like the real dashboard.
click to expand
Every step is logged as an event in the SQLite database. If the worker crashes, the dispatcher reclaims the task on the next tick and respawns. If the worker blocks, you get a notification (via Telegram, Discord, or any connected gateway platform) with the reason. You comment your answer, unblock the task, and the worker respawns with your feedback in its context.
These are not toy examples. Each blueprint is a real, deployable workflow that solves a genuine problem. Every one includes the full task graph, prompts, profile assignments, and pipeline structure — ready to paste into your Hermes Kanban board.
Turn passive content consumption into a structured, searchable knowledge base. Every article, video, or paper you encounter gets researched, summarized, cross-referenced with what you already know, and filed into your notes — automatically.
research, default (synthesizer), content (writer)
dir: pointing to your Obsidian vault
A multi-profile board that runs your life logistics 24/7: health monitoring, financial tracking, learning management, and home maintenance — each domain handled by a specialist agent with persistent memory. The orchestrator profile generates a weekly life briefing every Sunday.
research (health+finance), content (learning), dev (home maintenance), default (orchestrator)
dir: persistent directory for life ops data
For small agencies and consultancies managing multiple clients. Each client request flows through a standardized pipeline: intake → research → draft → human review → delivery. Clients are isolated as tenants on the same board. Human-in-the-loop at the review gate only — everything else is autonomous.
research (context), content (drafting/delivery), default (intake triage)
dir: per-client folder under Documents/Clients/
--tenant client-a, --tenant client-b for isolation
Orchestrate a full product launch across specialist profiles: market research → content production → technical QA → compliance review → timed launch execution. Each phase has parent-child dependencies. Human-in-the-loop at compliance and launch gates. The entire launch is reconstructable from the audit trail.
research (market+compliance), content (production), dev (QA+launch), default (orchestrator)
worktree: for code QA, dir: for content, scratch for research
--tenant client-a) to isolate client work on the same board. Use dir: workspaces pointing to per-client folders. The audit trail means you can show clients exactly what was done, when, and why. Human-in-the-loop at review gates ensures quality control before anything reaches the client.Hermes Agent is open-source, MIT-licensed, and runs on your machine. Install it, enable Kanban, and put your first task on the board today.