Inside Hermes Kanban: One Prompt Becomes a Task Graph, and a Custom Dashboard Adds the Human Gate

Hermes Agent ships a real multi-agent Kanban board with a SQLite-backed task graph, and the approve-once workflow in this video comes from a dashboard Julian Goldie built on top of it rather than from Hermes itself. The 34-second clip is a screen tour of a dark-themed workspace labeled "Agentic OS" in the top left of the sidebar. The left rail lists a workspace section (Mission Control, Paperclip, AI Agent Mastermind), an agents section (Claude, OpenClaw, Hermes, Gemini, Antigravity, Codex, Free Claude Code), and a self section (Pipeline, Goals, SEO, Video, Thumbnails, Studio, Notebook, Kanban, Ruflo Swarm, Journal, Memory, Build Guide). The visual language is consistent across every screen: near-black background, thin gray borders, a single accent color per page, roman-numeral breadcrumbs like "XIII / SELF / KANBAN" and "XV / SELF / MEMORY", and a "⌘K Command palette" plus an "ALL SYSTEMS" chip pinned to the top right. It reads like a terminal that grew a UI, and that restraint is most of why the demo feels credible.

What the three screens actually show

The first screen is a page titled "From Inbox to Shipped" with the subtitle "Capture an idea, agents classify, route and plan it, you approve once, a PM and subagents build it. Lives in your vault." Under it sits a single wide input reading "Drop an idea. A project, a thought, a link, anything. Agents take it from here." with a Capture button on the right. Four columns follow: Stage 01 Capture (Raw input to orchestration), Stage 02 Human Gate (Review checkpoint), Stage 03 Execute (PM and subagents build), Stage 04 Shipped and Filed. Capture holds cards like "Plan 'AutoResearchClaw vs. Ahrefs' comparison video" with a "Shape it" button, "Build a podcast clip generator", and "create a to do list app". Human Gate holds "AI cold email writer from LinkedIn" and "Podcast Clip Generator", each with a green Approve button and a red reject button beside it.

The second screen is the Kanban page. Its subtitle is "Hermes Agent multi-agent board. Drop a prompt into triage, watch the orchestrator decompose + assign." Controls across the top include a search field, an "All assignees" filter, "Show archived", "Refresh", and a "Dispatch now" button. Below that is a create row reading "New task title... (⌘+Enter to create)" with an assignee dropdown set to "julian" and a "Triage" checkbox. Six columns are visible: Triage, Todo, Ready, Running, Blocked, Done. Todo holds "Launch the AIPR explainer: explainer video, landing page, twitter thread, intro podcast". Blocked holds "Produce AIPR explainer video" and "Create a basic website". Done holds "Local SEO Playbook: dominate the map pack", "SEO Keyword Map: 50 high-intent clusters for AI automation", and "SEO Playbook: Programmatic SEO for AI tools at scale".

The third screen is Memory, subtitled "Search 1261 Omni memories + your Obsidian vault". It has Recent, Notes, Omni, and Graph tabs, a counter reading "1261 OMNI, 186 NOTES", and an animated force-directed graph labeled "MEMORY GALAXY, 256 stars, 1227 links". Node labels include Julian Goldie SEO, OpenClaw, Hermes Agent, Claude Code, and Decision Log. A status bar at the bottom shows an active file path: AGENTIC OS/PIPELINE/ITEMS/CREATE-A-NEW-SEO-WEBSITE-FOR-OPENCLAW-THAT-G-4-6.MD.

The Kanban board is real, and it is more rigorous than the demo suggests

Hermes Agent is an open source project from Nous Research at github.com/NousResearch/hermes-agent, MIT licensed. As of 2026-09-14 the GitHub API reports 245,327 stars and a latest release tagged v2026.9.11, named "Hermes Agent v0.21.2", published 2026-09-11. That is the same day this video was posted, which explains the "new" framing.

The official Kanban documentation describes a durable board backed by a SQLite database at ~/.hermes/kanban.db, with eight statuses: triage, todo, ready, running, blocked, review, done, and archived. The video's board shows six of those. Review and archived are filtered out of the visible columns, which matches the "Show archived" toggle sitting unchecked in the toolbar.

The engineering underneath is more careful than a Kanban skin implies. Tasks link parent to child through kanban_link, so completion metadata from a parent card flows forward into the context of its children instead of forcing a re-read of finished work. A dispatcher embedded in the gateway runs on a 60 second cycle by default, reclaims stale or crashed workers, and spawns profiles for tasks in ready. Configuration exposes max_in_progress, max_in_progress_per_profile, and a failure_limit that defaults to 2 consecutive failures before a task auto-blocks. Workers get one of three workspace kinds: an ephemeral scratch directory, a preserved git worktree, or a shared dir:<path> that requires an absolute path to avoid confused-deputy problems. None of that shows up in a 34 second clip, and it is the part that determines whether the board survives a week of real use.

The approve-once gate is custom, not a Hermes default

The video's central claim is that agents stop and wait: "Then it stops. It waits for you to approve before touching anything." That is what the Human Gate column shows, with explicit Approve buttons on two cards.

Hermes' own docs describe something different. Auto-decompose is on by default (kanban.auto_decompose: true), meaning the dispatcher automatically runs the decomposer on tasks landing in triage. Human-in-the-loop exists, but reactively. A worker calls kanban_block to surface a decision, or kanban_request_review parks a task in review for a named reviewer. There is no documented mandatory gate that holds every task before execution.

So the "From Inbox to Shipped" pipeline, with its Capture, Human Gate, Execute, Shipped and Filed stages, is a layer Goldie built. His own writeups confirm this pattern. His Agentic OS build guide describes a mission-control dashboard the reader assembles with AI help, and states that a ready-made template is available inside his paid AI Profit Boardroom community while the from-scratch path is free. The vault file paths visible on screen, agentic-os/pipeline/items/*.md, confirm the pipeline is markdown in an Obsidian vault rather than a Hermes feature. That is a sensible design. It just means you cannot install Hermes and get this screen.

What the agents produced, and what I could not verify

The demo opens a completed card, "SaaS pricing page design", tagged project 95% and pointing at agentic-os/pipeline/items/a-saas-pricing-page-with-three-tiers-und.md. The classification block reads "Routed to project, confidence 95%, tags: web, design, saas. A multi-step undertaking, drafting a plan for your review." A live iframe labeled "What the agents built" renders a pricing page for a fictional product called Lumen, headlined "Pricing that scales with your ambition." and scrolling to a "Questions, answered." FAQ block. A gallery view shows more finished artifacts with "Preview live" buttons: a 3D artist portfolio website, a retro neon space shooter game, an AI consulting landing page branded "Neural Edge" with a badge reading "Now booking Q1 2026 engagements", a habit tracker app, a meditation timer web app, and an SEO calculator tool.

Two of the Done cards are opened. "SEO Keyword Map: 50 high-intent clusters for AI automation" renders a table with columns for cluster, example keyword, monthly volume, difficulty, and intent. Visible rows include "AI Agents / ai agent for business / 12,100 / Medium / Commercial", "Workflow Automation / ai workflow automation tools / 8,400 / Medium / Commercial", "Prompt Engineering / prompt engineering best practices / 9,500 / Low / Informational", and "Chatbots / ai chatbot for customer... / 9,800 / High / Commercial". A second table adds star ratings and P1 through P4 priorities.

Those search volume numbers cite no source anywhere on screen. No Ahrefs, Semrush, or Google Keyword Planner attribution appears in either frame. The same keyword appears at two different volumes across the two tables in the same document, with "ai workflow automation tools" listed at 8,400 in both, while "prompt engineering best practices" shows 9,500 in the first table and 6,300 in the second. I could not verify any of these figures against a keyword data provider, and the internal inconsistency suggests the agent generated plausible-looking numbers rather than pulling them from a tool. If you plan content against output like this, check the volumes yourself before committing a quarter of writing to them.

Where the demo is thin

The Running column is empty in every frame that shows the board. Nothing is observed executing. Blocked contains two cards, including "Produce AIPR explainer video" and "Create a basic website", which is the honest failure mode of an autonomous board and is visible right there in the screenshot. The narration's "you do something else" promise is illustrated with finished artifacts and an empty Running lane, so the clip demonstrates the inbox and the outbox without showing the middle.

Every shipped artifact in the gallery is a self-contained front-end build or a markdown strategy document. A pricing page, a portfolio site, a browser game, a habit tracker, and keyword research all share one property: no external system to integrate with, no existing codebase to not break, and no correctness check harder than "does it render". That is the category where agent swarms currently do well. The video does not claim otherwise, and it also does not show the harder category.

The memory claim holds up better. "Everything gets logged into memory. So you never have to re-explain context again" maps to the Obsidian vault plus Omi captures, which is the stack Goldie documents in his own posts. The on-screen label reads "Omni" while the product is spelled Omi, so either the UI has a typo or it refers to something else. The underlying mechanic of agents reading and writing the same markdown vault is real, and it is the least flashy and most useful idea in the video.

Key Takeaways

• Hermes Agent is MIT licensed and free at github.com/NousResearch/hermes-agent. The GitHub API reported 245,327 stars on 2026-09-14, with release v2026.9.11 (Hermes Agent v0.21.2) published 2026-09-11, the same day this video went up.
• Hermes Kanban stores tasks in SQLite at ~/.hermes/kanban.db across eight statuses (triage, todo, ready, running, blocked, review, done, archived). The video's board shows six of them.
• The dispatcher runs every 60 seconds by default, reclaims crashed workers, and auto-blocks a task after 2 consecutive failures via the failure_limit setting.
• Auto-decompose is enabled by default in Hermes, so the "it stops and waits for your approval" behavior is not stock. The Capture, Human Gate, Execute, Shipped pipeline is Goldie's own layer, stored as markdown under agentic-os/pipeline/items/ in an Obsidian vault.
• The keyword volumes in the agent-generated SEO document (12,100 for "ai agent for business", 9,500 then 6,300 for "prompt engineering best practices") carry no source attribution and contradict each other across two tables in the same file.
• Everything visible in the Shipped gallery is a standalone front-end build or a strategy doc, which is the easiest category for agents. The Running column is empty in every frame.
• The sidebar treats Claude, OpenClaw, Hermes, Gemini, Antigravity, and Codex as interchangeable agent backends behind one dashboard, which is the actual architectural idea worth stealing from this clip.

Resources

• Hermes Agent on GitHub - MIT-licensed agent from Nous Research, the software the Kanban board belongs to.
• Hermes Kanban documentation - Official spec for the eight statuses, dispatcher settings, workspace kinds, and CLI commands.
• Hermes Agent site - Install instructions and the rest of the docs.
• Julian Goldie's Agentic OS build guide - How the dashboard in the video is assembled, and which parts are gated behind his paid community.
• Hermes Agent Kanban board writeup - Goldie's own longer explanation of the orchestrator and agent profiles.
• OpenClaw - Open source self-hosted agent listed in the sidebar, at github.com/openclaw/openclaw with 389,649 stars as of 2026-09-14.
• Ruflo - The multi-agent swarm harness behind the "Ruflo Swarm" sidebar entry, MIT licensed, 72,394 stars as of 2026-09-14.
• Obsidian - The local markdown vault used as the shared memory layer.
• Omi - The wearable capture device that feeds voice notes into the vault, labeled "Omni" in the UI.
• Google Antigravity - Google's agentic development platform, listed as an agent backend in the sidebar.