Skip to content

created: 2026-04-24 updated: 2026-04-24 tags: [source, video, youtube, coding-agent, pi] type: source url: https://www.youtube.com/watch?v=RjfbvDXpFls author: "AI Engineer" (Mario Zechner) published: 2026-04-17

Building pi in a World of Slop — Mario Zechner

Summary

Mario Zechner (badlogic) presents pi, his minimal open-source coding agent with 4 tools, sub-1000 token system prompt, and TypeScript extension system. Also discusses how AI agents are flooding OSS repos and his strategies to fight back.

Key Takeaways

Why Zechner Left Claude Code

  • Started using Claude Code in April 2025; initially simple and predictable
  • "Token madness" set in as the team grew — features he didn't need, more bugs
  • Context wasn't his context — Claude Code modifies system prompts and tool definitions behind the scenes, inserts system reminders that may or may not be relevant
  • Zero observability into what the agent is doing
  • Zero model choice (Anthropic only)
  • Shallow extensibility — hooks spawn new processes each time

Problems with Open Code

  • Prunes tool output after minimum token thresholds → "lobotomizes the model"
  • LSP server injects errors into edit tool results → confuses the model
  • Stores individual messages as separate JSON files on disk
  • Default security: server spins up with CORS headers allowing any website to access it

The Terminal Benchmark

Terminal — the most minimal coding agent harness possible: - Only gives the model a tool to send keystrokes to a tmux session and read output - No file tools, no sub-agents - Scores higher than native harnesses (Claude Code, Codex) on the leaderboard

Thesis: Coding agents are in the "poke around and find out" phase; current form is not final form.

pi: Self-Modifying Agent Core

  • 4 tools: Read, Edit, Bash, Mesh (that's it)
  • Sub-1000 token system prompt — models are reinforcement-trained to know what a coding agent is
  • TypeScript extension system — hot-reloadable, full control over tools, UI, events
  • YOLO security — no permission prompts by default; users can build their own permission gates
  • Skills support — added the industry-standard Skills format (markdown files)
  • Session management — tree view, fork/branch from any point in conversation

pi Extension Examples

  • Agent-to-agent chat rooms (Nico's Pi agents talking to each other)
  • NES game emulator inside the agent
  • Doom inside the agent
  • Git status in the UI
  • Context workflow extension (encode multi-step development pipelines)

OSS in the Age of "Clankers"

  • pi was adopted as the agentic core of OpenClaw → Mario's repo got flooded with AI-generated issues/PRs
  • Fightback strategies:
  • Auto-close PRs with a comment asking for human-written issues (clankers don't read the comment)
  • "Vouch" system — accounts that pass the test get whitelisted
  • Deprioritize issues from accounts that interacted with OpenClaw
  • Embed issues/PRs into 3D space to see clusters of garbage
  • "OSS vacation" — close the tracker whenever he wants

Act 3: "Slow the Fuck Down"

  • Agents compound "boooos" (errors) with serial learning, no bottlenecks, delayed pain
  • 10 agents × 2 humans = enterprise-grade complexity in 2 weeks
  • Agents learn complexity from the internet — 90% of code on the internet is "our old garbage"
  • Every agent decision is local, especially when the codebase exceeds context window
  • "SAS is dead, software solved in six months" — the irony
  • Agents don't learn the way humans learn — they have no pain mechanism
  • Humans feel pain → quit, blame, refactor; agents keep digging deeper

Practical Advice for Working with Agents

  • Scope — modularize codebase so agent can find everything it needs
  • Evaluation — give it a function to evaluate how well it did
  • Hill climbing — auto-research, boring stuff, let it wipe
  • Non-mission-critical — reproduction cases for user issues
  • Rubber duck — agent as sounding board when no human is available

Final Principles

  • Fewer features, but the ones that matter
  • Get the amount of generated code you need to review down
  • Non-critical code: wipe slop ahead
  • Critical code: read every line
  • If you do anything important, write it by hand — the friction builds understanding in your head
  • "Learn to say no — this is your most valuable capability at the moment"