All articles
The run log, 11 articles.
Back to School
Prompting couldn't teach a small local model to find its way out of a maze. Fine-tuning a model half its size on a gaming GPU did, and the thing that nearly stopped it was a single word.
Maze Escape
Three agents, a maze, a notebook each and a radio. I wanted to see whether they could help each other escape.
One click, one trace
Following a single browser click through two agents, three runtimes and an MCP hop, to see whether OpenTelemetry keeps it in one trace.
Lost in translation
What Jaeger and Pydantic Logfire made of the same coding-agent telemetry, and how LangSmith, Langfuse and Braintrust compare on paper.
One standard, three dialects
What Claude Code, Codex and Qwen Code send over OpenTelemetry.
What went over the wire
Moving the logging out of my agents and onto the wire.
Models & Monsters
The Goblin who filed feature requests: How unexpected AI behaviour shaped my multi-agent harness.
Can a small model be a good agent?
Twenty-eight models, 1,058 runs and a surprising result: small local models can use tools remarkably well—until the obvious next step fails.
Squeezing the juice
How many things can an 8GB card do at once? Fewer than it will let you ask for. In this post I benchmark parallelism in a local llm
What the model was thinking
Twenty times the tokens to reach the same answer.
What the model actually saw
Four documents, one summarisation, and an answer that took a while to look wrong.
