Software and AI Solution Architect at white duck, Rosenheim / Munich.
I ship software through coding agents. People own the specification, the checks and the acceptance, and the agents write the code. Rules belong in code, not in prompts.
What I build
- Flock: a contract-based agent system in Python. Agents declare typed inputs and outputs and coordinate through a blackboard instead of hand-wired graphs. Ships an adapter for Microsoft Foundry Hosted Agents.
- Tare: proves that your coding agent starts without your personal setup before you measure it. Works with Codex, Claude Code, Pi and Antigravity.
- Dojo: creates and evaluates agent skills by comparing baseline and candidate runs.
- Atlas: grow a knowledge vault by talking to your agents, and read it like a book. Live demo.
- tendril: an experiment with the Codex SDK. Agent threads are application state on a graph, and every change passes proof checks and human approval.
- wd-lecture-agents-2026: a four-hour hands-on workshop on Codex, Claude Code and GitHub Copilot (German).
Writing
Infinite Loop: interactive essays, each post its own small app. Latest: Fifty-One Hours and Who Decides Done?.
Lately
- Ran a coding agent on Flock for 51 hours: 88 merged pull requests, 3,289 backend tests on every run, and Codex as reviewer.
- Talks on agent orchestration at Mastering Microservices (2025) and Experts Live (2026).