Notes in May 2026

Today, I started trying out Pi, the minimal coding agent created by Mario Zechner and now developed at Earendil. It is known for its minimal harness, few tools, no built-in MCP support, etc., and high extensibility.

I haven’t settled on my perfect Pi setup yet, but I’ve already found a neat way to add MCP support with MCPorter, which lets me use Exa MCP for web search. I also discovered that defining a custom provider makes it easy to use my company’s model provider, which was originally designed for other agents.

Overall, Pi’s minimality and flexibility make everything clear and nothing magical, which feels very developer-friendly. I’m considering dropping Codex and Claude Code, and using Pi along with Amp as my main agents.

Pi

Here are some articles and videos I bookmarked about Pi:

• Mario Zechner’s post where Pi originated: What I learned building an opinionated and minimal coding agent
• Armin Ronacher’s introduction: Pi: The Minimal Agent Within OpenClaw
• Mario’s post announcing Pi and his move to Armin’s company Earendil: I’ve sold out; and meanwhile, Armin’s post: Mario and Earendil
• Armin’s post: Building Pi With Pi
• Videos on YouTube:
  • Claude Code is overkill - Pi is All you Need
  • A love letter to Pi | Lucas Meijer
  • I Hated Every Coding Agent, So I Built My Own — Mario Zechner (Pi)
  • Building pi in a World of Slop — Mario Zechner
  • Building Pi, and what makes self-modifying software so fascinating

ProgramBench¹. A new benchmark that evaluates whether language models can rebuild black-box software systems from scratch.

It was created by Meta, and co-first author John Yang is also the creator of SWE-bench.

Unlike previous benchmarks such as SWE-bench, which work with existing codebases, each ProgramBench task gives the model only a compiled binary and its documentation. The model must then implement a complete codebase that reproduces the original program’s behavior. During this process, the model has no internet access, and decompilation tools are banned.

This is a difficult benchmark. As of today (May 14), only one task had been fully solved by GPT-5.5², giving it a score of 0.5% across the 200 tasks.

ProgramBench leaderboard