Hermes Kanban
hermes-kanban is Hermes Agent's durable multi-agent coordination board. Instead of treating collaboration as an in-memory swarm of temporary subagents, it stores tasks, comments, dependencies, workspace assignments, and worker state in a persistent SQLite-backed board shared across Hermes profiles.
Core idea
Hermes Kanban turns multi-agent work into a database-backed queue and state machine.
The durable objects are:
- tasks with assignees and statuses
- links between tasks
- comment threads as handoff history
- workspaces where workers operate
- dispatcher claims and liveness signals
So every handoff becomes inspectable and resumable.
Why it exists
hermes-subagent-delegation is excellent for short-lived reasoning branches, but it breaks down when work must:
- survive restarts
- wait on human input
- move between named agents over time
- be audited later
- accumulate a long-lived comment trail
hermes-kanban is Hermes's answer to that durability problem.
Profile-centered collaboration
A Kanban worker is not an anonymous child summary. It is a full Hermes profile with its own memory, skills, configuration, and identity.
That means the board coordinates named long-lived agents, not just ephemeral branches of one parent prompt.
Dispatcher model
A dispatcher loop watches the board, promotes ready tasks, reclaims stale claims, and spawns the assigned worker profile when work is ready.
Because the dispatcher normally runs inside the hermes-gateway, the board inherits the same persistent daemon substrate that powers messaging and cron delivery.
Tool-mediated worker protocol
Hermes workers do not coordinate through shelling out to a CLI wrapper. They use a dedicated kanban_* toolset to read tasks, post comments, heartbeat, block, or complete work.
This matters because it makes the board a first-class agent substrate, not just a shell-level queue bolted onto the side.
Kanban versus delegation
The distinction from hermes-subagent-delegation is fundamental:
- delegation is a parent-owned synchronous fork/join call
- kanban is a durable peer-readable work queue
Delegation returns results into the parent context and disappears. Kanban leaves a durable artifact trail that any authorized profile or human can inspect later.
Kanban versus cron
The distinction from hermes-cron is also important:
- cron answers *when* an autonomous run should happen
- kanban answers *how* multiple named agents coordinate across time
A cron job can create or monitor board work, but the board itself is the coordination memory.
Architectural consequence
hermes-kanban shows Hermes moving beyond "agent with tools" into agent operations infrastructure.
It combines:
- durable state
- named worker identities
- explicit handoffs
- dependency tracking
- human intervention points
- resumability after crashes or blocking conditions
This makes it one of the clearest examples of Hermes treating multi-agent collaboration as an operating system problem instead of a prompt-engineering trick.
See Also
- hermes-agent
- hermes-gateway
- hermes-profiles
- hermes-cron
- hermes-subagent-delegation
- shared-memory-blackboard
- multi-agent-patterns-moc
- pattern-progressive-handoff
References
- Source:
C:\Users\executor\AppData\Local\hermes\hermes-agent\website\docs\user-guide\features\kanban.md - Source:
C:\Users\executor\AppData\Local\hermes\skills\autonomous-ai-agents\hermes-agent\SKILL.md