Skip to content

Pi Agent's Developer on Why Minimal Beats Maximal in AI Coding

A teal π symbol, a nod to the Pi coding agent

Pi is one of the smallest coding agents around — a minimalist terminal "harness" with a system prompt under 1,000 tokens and just four built-in tools (read, write, edit, bash) — and yet it keeps beating far heavier rivals like Claude Code and Codex in head-to-head comparisons. In this episode of the David Ondrej podcast, Armin Ronacher (the creator of Flask and Jinja2, who now works on Pi at Earendil together with Mario Zechner) explains why less tooling is winning, and walks through his own agentic engineering workflow.

Less Tooling, More Bash

Ronacher's explanation is disarmingly simple: modern models have become extremely good at using computers, and Pi basically just gives them bash. Where older harnesses tried to be helpful by wrapping every capability in bespoke tools, that has stopped paying off — even Codex largely relies on bash now (it calls rg to find files) and pipelines commands together to stay context-efficient rather than dragging everything into the prompt.

He argues the underlying reason is training data: reinforcement-learning setups reward the low-level approach, so the models are simply better at it. And the direction of travel is unmistakable — Claude Code now ships fewer tools than at its peak, and OpenCode 2 is built around plugins rather than a big tool list. Pi's extensibility, not its feature count, is what made it popular.

The Workflow and the Weird New Realities

The second half turns to practice. Ronacher's own setup is local-first: cloud agents sound appealing, but wiring up Tailscale and SSH makes the UX clunky for now. He also flags two consequences that show up everywhere once agents start writing code at scale — GitHub struggling to keep up with the sheer volume of commits, and the fact that "number of commits" is a terrible productivity metric when the honest measures (support tickets closed, real problems solved) are much harder to see.

Where It's Heading

There is genuine uncertainty in his answers too. Coding agents went from one approach for AI to the approach in remarkably little time, new harnesses appear almost weekly, and the model labs themselves may end up competing in this space. His own edge, he suggests, is not naval-gazing about Pi's popularity but taste and judgement about where the field should go next — the same instinct that built a minimalist harness in the first place (the one, incidentally, that powers OpenClaw).

The full hour goes much deeper than this — specifics on his own engineering setup, agent traffic in open source, and why the smallest prompt may be the strongest one. Well worth watching if you build with coding agents.

Sources