Hermes Agent v0.12.0 adds Kanban boards for multi-agent workspaces

• Hermes Agent v0.12.0 adds a Kanban workflow where specialized agents claim tasks from a shared board, run in parallel, and hand work off when blocked, according to NousResearch's launch thread.
• The official Kanban tutorial says the dashboard and CLI share one SQLite board at ~/.hermes/kanban.db, with six columns, dependency promotion, and per-run history.
• NousResearch's feature list says each task can carry its own profile, tools, and skills, while handoffs can move through a directory, a git worktree, or a scratch space.
• Two GitHub issues filed right after release exposed rough edges in that shared-state model: one said the board became profile-scoped instead of global across profiles in some cases, and another reported a worker crash around SQLite access, per issue #18442 and issue #18657.
• Early users were already treating the board as the missing control plane for model switching and parallel runs, from LLMpsycho's profile-switching note to Teknium's repost of a 96-agent run.

You can read the official tutorial, browse the Hermes Agent repo, and see the release land in the open with a shared-board bug and a worker DB handoff fix reference. The weirdly useful detail is that Hermes did not just bolt on a board UI. The docs describe a dispatcher, dependency promotion, metadata-rich run history, and a dashboard that mirrors the CLI instead of replacing it.

The core change is simple: Hermes now has a board where agent profiles pull assigned work instead of living in isolated terminal sessions. NousResearch framed it as multi-agent via Kanban, with one place to watch progress and unblock work.

That board is not just a visual wrapper around chat. NousResearch's feature list says each task runs on a specialized profile with its own tools, skills, and personality, which makes the board a scheduler for distinct worker configurations, not just a task list.

The Kanban tutorial lays out six columns:

1. Triage, for raw ideas that need specification.
2. Todo, for tasks waiting on dependencies or assignment.
3. Ready, for tasks assigned and waiting for a dispatcher tick.
4. In progress, for claimed work, optionally grouped by assignee.
5. Blocked, for tasks waiting on human input.
6. Done, for completed work.

The same tutorial says task dependencies are first-class. Parent tasks can gate child tasks, and finishing the parent automatically promotes dependents from Todo to Ready.

It also adds traceability that most agent demos skip:

• Run History records worker, duration, timestamp, outcome, and a handoff summary, per the tutorial.
• Metadata such as changed files or decisions can travel with the run and surface to downstream workers, again per the tutorial.
• The dashboard exposes filters, a "Lanes by profile" toggle, and a "Nudge dispatcher" button to trigger dispatch immediately, as shown in Teknium's screenshot and described in the tutorial.

The persistence model is doing a lot of the real work here. The tutorial says the web dashboard at 127.0.0.1:9119 and the CLI both operate on the same SQLite file, so the board becomes shared state for dispatch, status, and history.

According to NousResearch's feature list, Hermes pairs that shared board with several handoff mechanisms:

• shared directories
• git worktrees
• temporary scratch spaces
• comments on each task, readable by the next agent
• heartbeats and runtime caps for long-running jobs
• project isolation for multiple projects on one machine
• a no-double-claim guard when many agents race for the next task

That architecture also showed its seams immediately. In issue #18442, a user reported that the board was resolving relative to the active profile's HERMES_HOME, creating separate kanban.db files per profile instead of one shared board. In issue #18657, another user reported a Kanban worker crash that they traced to SQLite access during task completion; the issue now points to PR #18670, titled "Fix Kanban worker board DB handoff."

The first outside reactions focused less on the board itself and more on what persistent coordination unlocks. LLMpsycho said tasks survived switches between DeepSeek and Claude profiles, which matches the docs' claim that the board sits above any one active worker session.

The bigger flex came from Teknium's repost of a 96-agent run, which described 96 concurrent Hermes agents consuming 382,745,618 tokens across 171,136 API calls to deepseek-v4-pro. Even as a repost, it shows how Nous wants the feature read: not as a prettier todo list, but as a control surface for runs that would be painful to manage by hand.

A separate branch of the Hermes ecosystem is already leaning into that framing. Teknium's Hermes Workspace repost points to Hermes Workspace as a UI layer with a Conductor view for spawning and monitoring parallel agents, which makes the Kanban release look like the open source control plane catching up to the broader Hermes stack.