About
Who I am, what I work on, and what I write about.Who I am
I'm Magnus Rødseth, a developer in Oslo. I work as a full-stack developer at Capra Consulting, currently building Hei, huset! for Gjensidige: a conversational AI assistant that lets homeowners talk to their house and get maintenance plans and advice drawn from their own property documents and structured housing data.
Most of my day is agent orchestration in Python (LangGraph, LangChain, Azure OpenAI), streaming document extraction with RAG on PostgreSQL and pgvector, and the frontend that sits on top of it in React and TypeScript. The other half is dragging the team's way of working toward agents that can be trusted with real work: living AGENTS.md docs, domain-specific skills, and an agent-agnostic MCP setup.
I hold a Master's degree in Computer Science from the Norwegian University of Science and Technology, specialising in Software Systems. I'm an Anthropic Claude Certified Architect and a Board Member at Junior Consulting, where I was previously Partner and Head of Technology.
What I build on the side
I build and sell Eden Stack, a production-ready monorepo for AI-native SaaS. Smaller things end up on GitHub: most recently sparebank1-cli, a Rust CLI that gives an AI agent read access to my own bank data, which an agent wrote almost end to end while I was at work.
A full list of projects and work history lives on this site, and the tools I actually use every day are listed too, kept honest rather than aspirational.
What I write about
I write about agentic software development, mostly in Norwegian: how to prepare a repository for agents that run unattended, spec-driven development, context engineering, the Model Context Protocol, ReAct and planning loops, multi-agent orchestration, evaluating systems that are not deterministic, prompt injection and agent security, what the EU AI Act actually asks of developers, and where the token economics land.
Posts are experience reports, not documentation. They are dated and I do not revise them as models change, so read the date before you trust a benchmark or a price in one. Several first ran at Capra Consulting or Kode24; those say so at the top and link back to the original.
Everything is at /blog, and there is a feed if you would rather not check back.
For agents
This site publishes a machine-readable copy of itself. Start at llms.txt for an index, or append Accept: text/markdown to any page URL to get that page as Markdown. openapi.json describes every endpoint; the blog is available as JSON at /api/v1/posts. No key and no rate limit.