Agent Engineering
About
I build AI agents, test what works with my own experiments, and publish the results here, including the ones that failed.
I'm Daniel Huber. Agent Engineering is where I write down what I learn while building AI agents, and where I test claims about them instead of repeating them. It is for engineers who build these systems, not for readers of announcements.
What's here
- Experiments are the core of the site: my own measurements on questions people building agents face, such as whether a small model can guard an agent against prompt injection, or whether a self-improving loop learns anything new. Each states its method, how sure the result is, and what went wrong. Most are pre-registered: the pass rule is written down before the data exists. Negative results are published like positive ones.
- Topics are reference pages on the core patterns: tool use, memory, multi-agent orchestration, retrieval, MCP and A2A, evaluation, safety. They come with diagrams and code in Python (LangChain/LangGraph) and C# (Microsoft Agent Framework).
- Perspectives are analyses of where agent engineering is heading, written from the news of the past two weeks.
- Articles also include summaries of research papers and posts, always linked to the original.
- Foundations is an interactive guide to the machine learning underneath, and the glossary and resources collect terms and further reading.
Where AI is used, and where it isn't
A lot of this site is produced with AI, and every page says so where it applies:
- News feed. Every day a script collects posts from about a dozen company, practitioner and research feeds and a set of accounts on X. Claude selects the ones relevant to agent engineering and summarises each in a sentence. Nothing is written by hand.
- Perspectives. Every other week, Claude reads that period's news and anything I add to a reading list, and drafts at most one analysis. An editorial check rejects drafts that repeat the shape of earlier ones or lack sources, and a week without a real thesis produces no article. I review each draft before it is published. These pages say "AI-generated, human-reviewed".
- Summaries. Articles based on someone else's paper or post are AI-generated summaries; each names and links the original authors at the top.
- Experiments. The questions, designs and decisions are mine. Code and analysis are written with AI coding agents, and every number on an experiment page is checked against the raw results before it is published.
- The site itself is built with Claude Code.
The stack
The site is built with Astro and Tailwind CSS and deployed on Vercel. Content is MDX in Astro content collections; charts use Plotly and diagrams Mermaid. The automation runs on GitHub Actions with the Anthropic API. Experiments run on my own machines, from a small home server to a single NVIDIA DGX Spark.
Contact
Corrections, questions and ideas for experiments are welcome: danielhuber.dev@proton.me, or on LinkedIn.