From working with AI, from teaching, from the DGX Spark, from conferences.
Each of my six AI advisors sees only its own field, yet almost every important question involves several of them. So I gave them a shared council: first as a conversation format, then as a morning meeting that plans four hours of my day.
In one day I set up five AI advisors in Claude Code, and a sixth a few days later. The hard part was not writing their roles but working out what they should remember, who may read it and how much it may cost.
One config file, a gateway running around the clock on a DGX Spark, local models and Claude. I go through my OpenClaw configuration piece by piece (models, agents, sandbox, channels, skills) and four things I only learned about it once it was running.
Notes nobody reads are just a more expensive way to forget. In April I turned it around: I collect the sources, and AI writes and maintains the wiki. Six months later it has almost five hundred pages, and the most valuable thing about it is the mistakes.
The machine I run local models on shut itself down twice in one evening — no error message, not a single line in the log. In the end it was a known defect that the manufacturer acknowledges. More interesting than the defect, though, is the method that leads to it, and the three traps along the way that wait for anyone running their own AI infrastructure.
At TechEd 2026 — the largest IT conference in the Czech Republic and Slovakia — I'm giving two talks, and both follow the same idea: moving AI from the role of an assistant to that of a reliable partner. The first is spec-driven development — how to write specifications that AI actually understands, and how it really works today with tools like Google Antigravity or Gemini CLI with the Conductor plugin. The second goes the other way: instead of the cloud, a local agent on your own hardware (OpenClaw + DGX Spark), with no token limits and with data that never leaves. Together they form my picture of where AI-assisted development is heading in 2026 — and why it's worth mastering both directions.
Write to jan.tamas@tamasoft.net and I'll add you to the list.