Pi is the Claude Code Killer Nobody Saw Coming...
👉 Access our AI Architects course & join hundreds of serious AI builders in our community:
https://www.theaiautomators.com/?utm_...
🔗 Pi
Pi: https://pi.dev/
Package catalog: https://pi.dev/packages?type=extension
Why Mario Zechner built Pi: https://mariozechner.at/posts/2025-11...
Earendil acquires Pi (MIT core): https://earendil.com/posts/announcing...
🔗 Research & Benchmarks
SWE-agent (the harness-ablation paper): https://arxiv.org/abs/2405.15793
mini-SWE-agent (minimal harness): https://github.com/SWE-agent/mini-swe...
Terminal-Bench 2.1 leaderboard: https://www.tbench.ai/leaderboard/ter...
SWE-bench leaderboard: https://www.swebench.com/index.html
🔗More
Building pi in a World of Slop (Mario Zechner): • Building pi in a World of Slop — Mario Zec...
Most people treat the AI model as the whole product. But take the exact same model, wrap a different coding agent around it, and the result can change dramatically. Think of the model as an engine and the harness as everything else in the car - the tools it can reach for, what it's allowed to see, how it runs the loop turn after turn, and how it recovers when something breaks. That's what actually decides how far the raw power gets you.
In this video we break down the two competing philosophies for how a harness gets built. On one side, the batteries-included agents - the likes of Claude Code and Codex - ship with huge bundles of tools, long system prompts and an opinionated way of doing almost everything.
They're brilliant to pick up, but that convenience has a cost: they get bloated, they're a black box when something goes wrong, and they're generally happiest on expensive frontier models. On the other side is Pi, built by Mario Zechner - a deliberately stripped-down agent: a tiny, inspectable core of a few tools, an AGENTS.md file and plain-JSON sessions you can branch, that you point at any OpenAI-compatible endpoint, including a local model on your own machine.
What Pi leaves out is really the point. No MCP, no sub-agents, no permission prompts, no plan mode - you add those back deliberately, from a catalog of thousands of community packages or by getting Pi to write its own TypeScript extensions. And no security either, which is the catch: the moment you run it, sandboxing is on you (Docker), because nothing out of the box stops it running destructive commands.
So which harness should you actually use? For most people - especially anyone less technical - Claude Code and Codex are still the best pick, shipping updates continuously with real safeguards already in place. Pi starts to make sense for power users, dev teams, enterprise setups and anyone running local models, who need the control and are willing to own the sandboxing. There's no universal best coding agent for every model - the only real answer is to test the model-and-harness pair on your own work.
Timestamps:
00:00 The model isn't the whole product
00:16 Two philosophies
01:11 Same model, different harness
01:39 What is a coding harness?
02:09 What the research shows
03:07 Batteries-included up close
03:46 Pi up close
04:40 What Pi leaves out
05:35 Installing & sandboxing
06:31 The package catalog
07:41 Writing your own extensions
08:11 Which harness should you use?
09:25 No universal best