Video: "Claude Opus 4.8 Built an Agentic OS in 15 Minutes! (ultracode)" by Julian Goldie on YouTube.
What Julian asked Ultracode to build
The brief was a working Agent OS — a locally hosted dashboard connecting Claude, Hermes Agent, and OpenClaw into one interface with shared goals, a memory layer, and per-agent control panels. That's the same architecture Julian and others in the community have been assembling manually over the past few months, typically spending a full day on the initial setup.
With Ultracode on high-effort mode, he gave Claude a single goal prompt describing the dashboard structure he wanted. The model wrote its own orchestration script, spun up parallel sub-agents to handle the frontend, the backend routing, and the agent integration layers simultaneously, and compiled working code from the outputs.
What Ultracode actually contributed
The speed improvement comes from parallelism, not magic. Normally, building something like this sequentially means prompting for a component, reviewing it, prompting for the next, and repeating. Ultracode does the sequencing itself and runs stages concurrently where it can. The model handles the cognitive overhead of keeping multiple workstreams consistent — which is the part that usually forces you to move slowly.
In 15 minutes Julian had a working Next.js dashboard with a chat panel for each agent, a shared goals tracker, and a task board. It connected to Hermes via its API and provided basic routing between agent responses. Not polished, but functional — the kind of result you'd take as a starting point for further development rather than a finished product.
What the 15-minute figure doesn't include
Julian has been working with the Agent OS architecture daily for months. He knew exactly what prompt to give, what the output structure should look like, and how to validate that the code was correct. Someone starting from scratch would spend considerably more time writing a clear enough goal, understanding what Ultracode produced, and debugging the integration points.
The 15 minutes is also just the build session — it doesn't include the Hermes and OpenClaw setup that the dashboard connects to, or the Obsidian memory vault configuration. Those components were already in place. The fair comparison is: Ultracode dramatically reduced the time to build the dashboard itself, not the entire stack from zero.
What this pattern means for businesses that want AI agents
The interesting implication is that the gap between "wanting an AI agent dashboard" and "having one" has compressed significantly. Twelve months ago, this was a developer project measured in days. Now it's a Claude Code session measured in minutes, assuming you have a clear specification and the supporting tools already configured.
That said, a 15-minute build is a prototype. Production use — stable, maintainable, connected to your real business data and running reliably — still requires proper development time. What's changed is how quickly you can get to something worth showing, which is a real benefit if you're evaluating whether an AI agent stack makes sense for your operations.
Where this connects to NordSys
We build and configure AI agent stacks for clients — the kind of setup where Hermes, Claude Code, and related tools are properly wired together for your specific workflow rather than assembled from YouTube tutorials and held together with hope. If you want the Agent OS architecture without the build time, that's exactly what we do.
See our AI Agents service →