HomeAboutLabInsightsContact

Lab

What I build myself.

I'm building myself a small agent organisation: AI agents develop software and content towards the goals I set, and I lead the whole as the client and decision-maker. What matters isn't the technology but the leadership model. I've designed, built, tested and, when needed, discarded parts of the model myself.

The system in four layers

  1. Digital mind

    A personal AI workspace: memory, principles, working instructions, content work and goals. It's built so that language models and technology providers can be swapped.

  2. Product owner agent

    Orchestrates software projects: turns goals into work cycles, directs the agents and reports the situation to me. It is the only managed channel between the personal workspace and the agents.

  3. Isolated development environment

    The agents' workshop. Agents can install and build freely, but have no access to personal data. A human can take the controls at any moment.

    Isolation boundary

Own server

The foundation everything runs on: encrypted disk, private network, backups and controlled releases.

How the work flows

  1. 01 Human

    Goal

    I write down a goal and its rationale for the project, for example "improve the reading experience of the article page".

  2. 02 Agent

    Proposal

    The product owner proposes a work cycle: what gets done, what I will see at the end and why this matters most right now. Work starts only after my approval.

  3. 03 Agent

    Work cycle

    Up to about 8 hours. An implementer agent does the work and a separate reviewer agent checks the code and screenshots. Status reaches my phone every four hours, and a single word stops the work.

  4. 04 Agent

    Review

    At the end of the cycle I get a preview link, screenshots, the decisions made, open questions and a proposal for the next cycle.

  5. 05 Human

    Decision

    Only my approval takes changes to production. A rejection starts a fix within the same cycle.

Philosophy

  1. Humans decide direction, AI decides execution

    Delegation happens at the level of goals, not task lists. I approve the outcome, not every intermediate step.

  2. Trust is built with structure, not instructions

    Isolation, permissions, budgets and approval points live in the system, not just in the agent's instructions.

  3. Privacy first

    Personal memory is kept apart from agents that handle untrusted content from the web.

  4. Only results count

    Benefit is measured in finished results and time saved. If a process produces more reports than results, it goes.

  5. Tool independence

    My own principles, memory and working instructions live in my own library. Models and tools can be replaced without losing the know-how.

  6. Try, measure and discard boldly

    My own solution is always compared with the simplest alternative. If it doesn't win, it goes.