04 / Agents · Current workflow
Hermes + KanbanA video studio on one home PC.
I customised Hermes for my own video studio. Agents and scripts on one home computer prepare videos for six shows across five accounts. I approve finished episodes and publication.
CURRENT OPERATION
A team of agents on one gaming GPU.
I wanted to find out how much work a team of agents could handle on a normal gaming GPU without the cloud. The original Hermes supports agents and Kanban, but long tasks got stuck, cards overwrote each other's outputs, and the model spent time on steps a script could handle.
The entire operation runs locally: one language model, local image, video, voice and music models, custom web search and a source database. Scripts and cron handle repeated steps; the model makes decisions. Every production round goes through checks, and approval and publishing remain my responsibility.
- 01
Plan
Cron opens a new production round for a show. The lead role breaks the work into cards with clear outputs.
- 02
Create and verify
The script is grounded in saved sources. Citation checks, repetition checks and a visual gate stop weak results.
- 03
Publish and measure
Render and schedule publication. Metrics feed back into the next scripts.
IN OPERATION / OCTOBER 2026
How I produce videos on one computer.
My own system for preparing videos: from sources and scripts through visuals and audio to review and a publishing proposal. The models run on my computer. I personally approve finished episodes and publication.
ONE VIDEO PRODUCTION CYCLE
- 01Topic and sources
- 02Script
- 03Citation check
- 04Voice, shots, render
- 05Visual quality gate
- 06Publication
- 07Metrics
Metrics feed into future scripts. I personally approve episodes and publication.
The first three seconds matter.
TikTok retention curves showed that only 17–22% of viewers were still watching at the third second. The opening scene now ends with a question, and the pipeline checks for repeated scenes.
Writing is easy. Verification is harder.
The most common reason for rejecting a script was whether its citations actually supported its claims. The pipeline therefore keeps its own source repository and checks citations before rendering.
Less time waiting for the model.
Structured cards, concise instructions and scripts replace repeated model calls. Comparing the same time window on 16 and 18 Oct 2026: median model calls per card fell from 5.5 to 3.0, time spent with lost cache from 61% to 14%, and model reloads from 29 to 13. This compares one day with another; it is not a controlled benchmark.
BROWSER SIMULATION
Work has a flow. And clear rules.
Try selected workflows directly in your browser.
This is a separate simulation using fictional data.
HERMES + KANBAN / INTERACTIVE DEMOSIMULATION · NO SIGN-IN
Loading the demo… A text overview is available below.
What you can try in this demo
- Kanban — Take a sample card through execution and review.
- Try a conflict — Two cards claiming the same output, or inconsistent task instructions.
- What I changed — My extensions alongside the original Hermes.
- A video production round — In Video round, select Next step to follow the brief, a rejected citation, its correction and video preparation. Finish with Approve (Štěpán). In batches and Image by image change the preparation mode and restart the round; Start a card manually demonstrates waiting during concurrent work. Nothing is published. Counts and delays are illustrative.
TECHNICAL DECISIONS
The model decides. Scripts repeat. A person approves.
One card, one model
The RTX 4060 Ti 16 GB holds only one large model at a time. The language model and image or video models take turns; a watchdog warns when two GPU jobs run at once.
Task instructions as data
Each card carries the project root, owned outputs and dependencies. Before execution, a validator rejects overlapping outputs, paths outside the project and dependency cycles.
Decisions based on data
Metrics from four platforms are fetched three times a day. They showed that most TikTok viewers leave by the third second, so the opening scene of each video was revised.
Short instructions for routine steps
A call that only selects the next command runs without reasoning and with a fixed response format: the same answer in 0.9 s instead of 16.6 s. Longer reasoning is reserved for content and exceptions.
An MoE model instead of a dense model
A Kanban step took 55–60 s with Qwen3.6-35B-A3B and 3.5–4 min with a dense 27B model. On a 16 GB card, time to an accepted result matters more than a leaderboard.
Status without a model
The plugin answers status and metrics queries directly from the records, without calling a language model.
Move repeated steps into scripts
Cron opens rounds, collects metrics, makes backups and tidies the board without an agent. The model handles exceptions and content.
Watchdogs instead of trust
Concurrency and round watchdogs restore a manually bypassed state and report it on Telegram. Every recurring error gets a test.
One card at a time preserves the cache
A manual start once launched five cards at once, and workers wiped each other's model cache. Now only the gateway starts cards, a concurrency watchdog checks them, and images are generated in batches.
My contribution
My own operation built on open-source Hermes: nine roles, over 25 cron jobs without model calls, a card contract, quality checks, metrics and watchdogs. I develop and review it with Claude Code, with some preparation done using Codex.
Project and demo status
In October 2026, the system prepares and schedules videos for six shows daily. On 17 October 2026, the full video pipeline test suite had over 3,190 passing tests. Comparing two operating days shows fewer model calls and less lost cache; there is no controlled benchmark against an unmodified default Hermes yet.
MY OPEN SOURCE EXTENSIONS
Hermes: my workflow extensions.
I extend the open source Hermes framework for my own work with agents and Kanban tasks. I focus on clear deliverables, preventing tasks from overwriting each other’s files, and tracking who launched the work.
Structured task card
Project root, declared files, required skills and dependencies as data that can be validated.
Conflicts are caught early
A local validator checks paths, overlapping outputs, inconsistent totals and dependency cycles.
Traceable execution
Recording who launched the work and translating paths between the Windows host and Docker container.
Built on NousResearch/hermes-agent ↗. These additions are documented in local source changes. I did not create the original framework, and this is not a new measurement of the running service.
Declared dependencies are checked for cycles; they do not enforce task completion on their own. Whether a structured brief is mandatory depends on the configuration.
EVIDENCE AND LIMITS
The foundation is NousResearch/hermes-agent; the video pipeline builds on the open-source MoneyPrinterTurbo. Figures come from local configuration, operating logs and measurements from 18 September to 18 October 2026. This website shows neither account names nor credentials. The board below is synthetic and does not control the running Hermes.