Mario Zechner's Minimal Coding Agent

Table of content

Mario Zechner created libGDX, a Java game framework that won a Duke’s Choice Award. He wrote Beginning Android Games and spent 20+ years in R&D and game development. Now he builds AI coding tools.

His pi-coding-agent strips everything out: 4 tools, no permission prompts, no sub-agents, no MCP. It benchmarks competitively against tools with far more code.

Background

Period	Work
2004-2010	Computer graphics research, PhD
2010-2015	libGDX development, Beginning Android Games
2015-2024	R&D engineering, various companies
2024-present	AI agent tooling (pi-mono)

Zechner lives in Austria and maintains an active blog documenting his experiments with AI coding tools.

Why he built it

Zechner got frustrated with existing agents injecting context “behind your back that isn’t even surfaced in the UI.” He wanted:

Exact control over what enters the model’s context
Full observability into every agent interaction
Clean session format for post-processing and sharing

Most coding agents fail on all three. They hide system prompts, spawn invisible sub-agents, store sessions in opaque formats.

Four tools

Pi uses four tools:

read   — Read file contents
write  — Write file contents
edit   — Edit file sections
bash   — Execute shell commands

No specialized file operations. The model uses bash for ls, grep, find. Zechner’s take: frontier models have been “RL-trained up the wazoo” and already know what a coding agent needs. Adding more tools wastes context.

His system prompt is under 1,000 tokens. Claude Code’s is much longer.

YOLO mode

Pi runs without permission prompts:

No pre-checking bash commands
No confirmation for file operations
Full filesystem access

Zechner calls security prompts “security theater.” Once an agent can write and execute code, asking permission for individual operations does nothing. Either you trust the agent or you don’t.

# Install and run
npm install -g @mariozechner/pi-coding-agent
pi

What pi excludes

Feature	Why
Built-in to-dos	“To-do lists generally confuse models more than they help”
Plan mode	Use PLAN.md files instead
MCP servers	Token overhead. Playwright MCP uses 7-9% of context before you start
Sub-agents	“Using a sub-agent mid-session is a sign you didn’t plan ahead”
Background bash	Use tmux for better observability
Permission system	Security theater

Same logic for each: keep context clean, maintain observability, avoid hidden state.

Session management

Pi lets you branch and resume conversations at any point:

# View session tree
/sessions

# Resume from a specific point
/resume <session-id>

# Branch from current state
/branch

Sessions export to HTML. The /share command uploads to secret GitHub gists. Boris Cherny’s parallel workflow uses git worktrees for similar isolation.

AGENTS.md and skills

Project instructions load from AGENTS.md files:

# AGENTS.md

## Project Context
This is a TypeScript monorepo using pnpm workspaces.

## Conventions
- Use `pnpm` not `npm`
- Tests live in `__tests__` directories
- Run `pnpm check` before committing

AGENTS.md loads hierarchically: global settings, then project overrides. No complex system prompt engineering needed. Compare this to Jesse Vincent’s Superpowers which uses a similar hierarchical skill system.

Skills are on-demand instruction files:

# Load a skill
/skill typescript-strict

# Skills are just markdown files
~/.pi/skills/typescript-strict.md

Multi-provider support

Pi supports 15+ LLM providers:

Anthropic, OpenAI, Google, Mistral
Groq, Cerebras, xAI, OpenRouter
Amazon Bedrock, Ollama
OAuth for Copilot, Gemini CLI, Claude

Switch models mid-conversation:

/model claude-sonnet-4
# or Ctrl+P to cycle

Context transfers when you switch. The agent summarizes and continues with the new model.

The TUI

Most terminal UIs flicker. Pi uses retained-mode TUI with differential rendering, so only changed portions redraw. Zechner built @mariozechner/pi-tui to solve this.

The pi-mono toolkit:

Package	Purpose
pi-ai	Multi-provider LLM interface
pi-agent-core	Agent runtime with tool calling
pi-coding-agent	The CLI
pi-tui	Flicker-free terminal UI
pi-web-ui	Web components for chat
pi-mom	Slack bot
pi-pods	vLLM deployment

VibeTunnel

VibeTunnel turns any browser into a terminal:

vibetunnel start
# Access from phone, tablet, anywhere
# https://your-tunnel.vibetunnel.io

The terminal session persists on your machine while you connect from anywhere.

Takeaways

Principle	How
Minimal toolset	4 tools handle everything
Transparent context	Under 1000 token system prompt
External state	Plans and todos in version-controlled files
No permission theater	Trust the agent or don’t use it
Session branching	Resume and fork at any point
Provider agnostic	Switch models mid-conversation

Links

Next: Thorsten Ball’s Agentic Coding Vision