Ali Raza's Picture

Hey, I am Ali.

I build software in Istanbul, and I write about what I actually learned building it.

The cheapest harness depends on the model

Runta and Baseten just published a huge harness leaderboard. The interesting result isn't a winner. It's that the cheap option for one model is expensive for another, and most of the spend goes to confident failures.

The other robot already knew

Mnemosyne was a notebook in one agent's pocket. Hindsight is a cabinet both of them can open. I set that up. The part that still isn't done is the funny part.

My agent fleet setup, with Buzz as the relay

I got one Codex agent talking through Buzz. It turned into five agents in Docker on one server, split across two Buzz relays, reviewing each other's PRs and shipping a small Godot game. Here's what it took, and what broke.

Harness engineering: the layer everyone's building and nobody's naming

Claude Code, Codex, LangGraph, CrewAI, Pi — they all have different answers to the same question: what happens to your agent's knowledge when the session ends? Here's why that question matters, how each system answers it, and why I ended up building on top of one that gets it right.

DSPy: The Framework for Programming LLMs (Not Prompting Them)

After two years of building with DSPy, I wrote down everything I learned. Here's what changed about how I work with LLMs.