← Back to library

Transcript: Hermes Agent Kanban Feature — Multi-Agent Content Pipeline

Watch video

Derek Cheung | AI Agents Automation10:49Transcript ✅Added May 28, 1:14 am GMT+8

Source video ID: 2oKmF--xJAI

Transcript

  • 0:00 — Hermes agent just shipped a Kanban feature and it’s a big one. Up until now, Hermes ran one agent at a time. You’d prompt it, it’d do the work, done. That’s fine for simple tasks, but Kanban changes what’s possible. You can now run multiple agents in parallel, each with its own task, its own tools, its own context, coordinating through a shared board. That unlocks a whole new category of automation, complex multi-step pipelines that actually run themselves. To show it off, we’re building an automated AI content pipeline. A
  • 0:32 — research agent finds trending YouTube topics, a script agent writes the video script, an X optimizer agent rewrites it for maximum reach using the open-sourced X algorithm. Everything saves to Supabase. The whole thing kicks off from a single message and by the end, we’ll have it running on a weekly cron. Here’s what we’re covering. Hermes agent on Railway. A multi-agent Kanban team. A Supabase schema built by asking. Specialized agents, research, script,
  • 1:02 — and X optimizer. A full pipeline run and a weekly cron to automate the whole thing. The Kanban feature is what makes this possible. It’s in latest version of Hermes. Step one, get Hermes running using the Railway template. See the description for the link. Click deploy, set your Deep Seek API key, and Railway will boot Hermes in about 2 minutes. I’ve linked previous videos in the description if you need a setup walk-through. Here is what is new. The Kanban tab in the left sidebar. This is a new feature in latest version. Step
  • 1:34 — one is done. The main idea behind Kanban is that instead of one agent trying to do everything, you break the project into tasks and assign each to a specialist. Research to one agent, scripting to another, optimization to a third. Each one focused, each one running in parallel with the Kanban board being the communication method. Agents pick up work automatically and update their status as they go. The board is the shared scoreboard the whole team reads from, so nothing gets
  • 2:05 — dropped, nothing runs twice, and you can see exactly where everything stands at any moment. Now, let’s build the team. I’m communicating to my Hermes agent through Telegram. I’m building a content creation pipeline. Create four Kanban tasks for these agents. One, research agent. Find trending YouTube topics using Tiny Fish web search and save results to Superbase. Two, script agent. Turn a research topic into a complete video script. Three, X optimizer agent. Apply X
  • 2:36 — algorithm virality signals to generate an optimized post and thread. Four, storage agent. Coordinate Superbase rights across all stages. Assign each to its own profile and set them as to do. Watch the Kanban board. Four cards appearing in to do. All four running. Own process, own tools, own context. They don’t share state. They coordinate through the board and Superbase. That’s the team. Now, they need a database, and that’s Superbase.
  • 3:06 — Step three, the content pipeline database. The agents need somewhere to write their output. Research topics, script drafts, X posts, pipeline state. Four tables in Superbase. And if you’re not on a paid plan, you won’t need one. The free tier is permanent, $0 per month, dedicated Postgres, 500 megabyte storage. For a project like this, you won’t come close to the ceiling. All of it built by telling Hermes what we need. To install the Superbase agent skill, we’re going to type into our Hermes
  • 3:37 — agent the following. Install the Superbase agent skill from the URL. My project URL is the following, and my service role key is this. Both are in Superbase settings. Hermes installs the skill and confirms the connection. Now, the schema. The exact prompt is in the repo link in the description. Copy it directly. Here’s the gist of what I asked for. Create four tables: topics, scripts, X posts, and pipeline runs with foreign keys linking each stage, JSONB columns
  • 4:08 — for the X algorithm signal data, and RLS policies so each agent writes only to the tables it owns. Watch the Supabase dashboard. Tables appearing: topics, scripts, X posts, pipeline runs. Schema exactly as specified. RLS policies added automatically. The signals applied column in X posts table is JSONB. That’s where the X optimizer will store the specific algorithm weights it applied. We’ll see what that looks like in step six.
  • 4:39 — Four tables, all wired together with foreign keys, all secured with RLS. Built by asking, no SQL editor. That’s Supabase ready. Halfway through the steps. The team is running. The database is ready. Now we configure what each agent actually does. Step four, the research agent. Its job is to find trending YouTube topics using Tiny Fish and rate each one for content potential. Tiny Fish is a web infrastructure platform built specifically for AI agents. Search, fetch, browser, and
  • 5:11 — agent under one API key. And here’s the part worth highlighting. Web search and fetch are completely free. Now install the Tiny Fish skill into Hermes. Same pattern as the Supabase skill in step three. Please install this Tiny Fish skill. And paste in your Tiny Fish API key when prompted. The skill teaches Hermes when to reach for Tiny Fish search versus fetch and exactly how to call each endpoint. Now configure the research agent. Configure the research agent.
  • 5:41 — Tiny Fish access. Task: Find five trending AI automation topics for non-technical audiences. Score each 1 to 10 for virality, save to the topics table, mark done when finished. Here’s what this looks like running in Hermes agent in Telegram. The research agent is running. Tiny fish is running search queries, results flowing back. And over in Superbase, the topics table is filling. 15 topics in Superbase. Top result, AI
  • 6:14 — agents for beginners. No code build tutorials 2026. Trending score of 95. The agent pulled that from a live tiny fish search, not guessing, scoring based on what actually came back. Step five, the script agent. It reads the research agent’s output, picks the top-rated topic, and writes a full video script. Here’s the prompt to create this agent. Script agent, pick the top virality potential topic from Superbase. Write a full script. Cold open, seven steps,
  • 6:45 — payoff, friction, outro. Save it as a draft, mark the card done. Let’s go ahead and prompt Hermes. Script agent is now running. It pulled the top topic. AI agents for beginners. No code build tutorials 2026 with a trending score of 95. Now it’s writing. The draft is going into Superbase in real time, and the full script column is filling as the agent writes. Script agent done.
  • 7:15 — Step six, X first open source their recommendation algorithm in April 2023, then released a Grok-based rewrite, Phoenix, in January 2026. The 2023 engagement weights are public in the code. Replies are worth 27 X a like. When the author replies back to a commenter, 150 X. External links have been suppressed since 2023. X confirmed it publicly in 2024. An early velocity engagement in the first 30 to 60 minutes is one of the
  • 7:47 — strongest ranking signals. Here’s the prompt for this agent. X optimizer agent. Read the latest Superbase script draft. Rewrite as a post and thread. Engineer for replies. 27 X alike. No external links in the root post. Target five plus replies in the first 15 minutes. Post at 7:00 to 9:00 a.m. or 6:00 to 8:00 p.m. Score 1 to 100. Save to X post with signals underscore applied JSON B. Mark
  • 8:17 — done. Let’s go ahead and prompt Hermes. The X optimizer agent is now running. It has pulled the script from the table and created an X thread post. The Superbase X post table has been updated with a new row. And the main post and thread columns are now filled. The X optimizer agent is finished. Step seven. The individual agents work. Now, let’s run the full pipeline from a
  • 8:47 — single prompt and watch the Kanban board coordinate everything automatically. Run the full content pipeline. Research five new AI automation topics using tiny fish. Pick the top two by virality potential. Write a script for each. Optimize both for X using algorithm signals and save all outputs to Superbase. Coordinate via Kanban. Agents should not start downstream work until upstream tasks complete. Let’s go ahead and prompt Hermes.
  • 9:25 — The X optimizer agent is now running. It has pulled the script from the table and created an X thread post. The Superbase X post table has been updated with a new row. And the main post and thread columns are now filled. The X optimizer agent is finished. Four tables filled. Two research to X optimize content packages. The thread format puts the link in the first reply to account for the algorithm’s link penalty. The agent knows that because it read the source code. Now, here’s the bonus I mentioned
  • 9:55 — earlier in the video. Tell Hermes, “Set up a weekly content pipeline run every Monday at 7:00 a.m.” Hermes creates a cron triggered orchestrator. Every Monday, the agent team runs, fills Superbase, and goes back to sleep. You open the database Monday morning, and your content week is planned. No human in the loop. That’s what the Kanban feature enables. Not one agent doing everything. A coordinated team that replaces the project manager.
  • 10:25 — Hermes Kanban. Four agents, one message. Tiny Fish researches, Superbase stores, the X algorithm optimizes, and a cron job runs it every week. Each agent is excellent at one thing. That’s the point. What would you build with a setup like this? Newsletter drafts, LinkedIn posts, product descriptions? Let me know in the comments. Subscribe if you want to see what the team builds next.