Author
Marco Patzelt
Software Engineer · AI Harness Layer
Software engineer building agentic projects — AI agents, tool design, MCP integrations, and multi-agent systems in production. Here I write about what actually works.
Articles by Marco
All articles
Code w/ Claude London 2026: Managed Agents, Memory + Dreaming, Evals
My notes from Anthropic's Code w/ Claude London 2026. The stuff that actually stuck a week later: managed agents, memory and dreaming, agent decomposition, and the eval discipline I'm now wiring into my own production stack.
Claude Opus 4.8 /ultracode: I Ran a 200-Agent Swarm
A hands-on, honest review of Claude Code's new multi-agent orchestration. What dynamic workflows and /ultracode actually do, where they bite, and whether the token usage is worth it.
Brunnfeld: 1000 LLM Agents, Zero Instructions
Brunnfeld: up to 1000 LLM agents run a medieval village economy with zero behavioral instructions. Two Reddit posts went viral, getting 55k and 58k views.
Karpathy's autoresearch: Why Constraints Enable Autonomy
Karpathy's autoresearch runs 126 ML experiments overnight without stalling. The seven design patterns behind it reveal how bounded autonomy actually works.
RAG vs Context Engineering: When Retrieval Adds Overhead
RAG was built for 4K-token context windows. With 200K+ tokens available now, context engineering gives agents full data visibility without the retrieval stack.
Context Engineering for AI Agents: Why Bare Metal Wins
Frameworks need .md files because agents don't understand them. Bare metal uses training data the agent already knows. Less abstraction, better agents.
Get in touch
Let's connect.
If you're working on something serious in agentic infrastructure — tool design, harness engineering, orchestration loops — drop me a line.