← back
Building pi in a World of Slop — Mario Zechner
Takeaway
Coding-agent harnesses should be tiny, observable, and self-modifiable so the model — already trained as a coding agent — owns its own context.
Summary
- Mario built 'pi', a minimal/extensible coding agent harness, after frustrations with Claude Code's opaque context manipulation (system prompts, injected reminders, zero observability/model choice) and open-code's behaviors like tool-output pruning and LSP injection on edits.
- Cites Terminal-Bench leaderboard where a minimal tmux-keystroke harness outperforms native harnesses, arguing 'self-modifying malleable agents' beat feature-heavy ones.
- Pi ships four packages (AI abstraction, agent core loop, TUI framework, coding agent) with just four tools (read/edit/bash/etc.) and a tiny system prompt; models are already RL-trained as coding agents and don't need bloat.
- Extensions are TypeScript modules hot-reloadable in-session distributed via npm/GitHub; agent can write its own extensions from shipped docs/examples (skills as markdown, custom compaction, providers, hooks).
- Pi is YOLO by default — security via extension rope rather than per-call approval dialogs.
coding-agentsclaude-codeharness
Original description
All I wanted was a shitty coding agent that is truly mine. And I’d have loved to just tell you why and how I built pi. But then Peter decided to make it the agentic core of OpenClaw. And now pi is collateral. So yes, this is a talk about pi. But it is also a talk about how agents are destroying OSS, how I deal with that, and a plea to slow the fuck down. https://x.com/badlogicgames https://github.com/badlogic https://www.linkedin.com/in/mariozechner/ Timestamps 0:00 – Intro and motivation for building pi 0:29 – Act 1: Building pi and the frustration with existing agent harnesses 1:56 – Why current context management in tools like Cloud Code and Open Code fails 4:44 – The importance of minimal harnesses and the "Terminal" benchmark 5:35 – Introducing pi: A self-modifying, extensible agent core 7:27 – The "YOLO" security philosophy and extensibility through TypeScript 9:03 – Examples of pi extensions (chat rooms, NES, Doom) 10:46 – Act 2: OSS in the age of "clankers" and how to fight them 12:03 – Act 3: A plea to slow down and stop the "slop" in software development 13:58 – How agents create "enterprise-grade complexity" and why humans are still the bottleneck 16:12 – Practical advice: How to effectively integrate agents into your workflow