I Hated Every Coding Agent, So I Built My Own — Mario Zechner (Pi)
Visit link →
Game development veteran, creator of libGDX, and 17-year open-source contributor Mario Zechner tells the story of how he ended up building pi, his own minimal, opinionated terminal coding agent.
It started in April 2025 when Peter Steinberger and Armin Ronacher (Flask, Sentry) dragged him into an overnight AI hackathon. Within weeks, Mario was hooked on Claude Code — until he wasn't. There was feature bloat, hidden context injection that changed daily, the infamous terminal flicker, and zero extensibility for power users.
He then surveyed the alternatives — Codex CLI, Amp, OpenCode... Eventually, he came across Terminus — an agent that gives the model nothing but a tmux session and raw keystrokes. If that's enough for the model to perform, what are all those extra features actually doing?
Mario's thesis: we're still in the "messing around and finding out" stage, and coding agents need to become more malleable so developers can experiment faster.
Pi is his answer: four tools (read, write, edit, bash), the shortest system prompt of any major agent, tree-structured sessions, full cost tracking, hot-reloading TypeScript extensions, and nothing injected behind your back. No MCP, no sub-agents, no plan mode — but all of it buildable in minutes through extensions.
The community has already shipped pi-annotate (visual frontend feedback), pi-messenger (a multi-agent chatroom), and someone even got Doom running. On TerminalBench, pi with Claude Opus 4.5 landed right behind Terminus — before it even had compaction.
🔗 LINKS & RESOURCES
pi coding agent: https://pi.dev
Mario Zechner: https://mariozechner.at
Peter Steinberger / OpenClaw: https://github.com/steipete
Armin Ronacher: https://lucumr.pocoo.org
Claude Code: https://docs.anthropic.com/en/docs/cl...
Aider: https://aider.chat
OpenCode: https://github.com/anthropics/opencode
Amp (Sourcegraph): https://sourcegraph.com/amp
TerminalBench: https://terminalbench.com
Ghostty: https://ghostty.org
Vouch: https://github.com/mitchellh/vouch
libGDX: https://libgdx.com
AI Engineer London is a community meetup for engineers and founders building with AI, covering everything from agent frameworks and RAG pipelines to LLMs in production. Each event features technical talks, live demos, and hands-on networking. This talk was recorded at AI Engineer London #10, hosted by Tessl, in collaboration with AI Engineer London.
AI ENGINEER LONDON
📅 Events: https://lu.ma/aiengineerlondon
💼 LinkedIn: / ai-engineer-london-meetup
📚 MASTRA RESOURCES
Mastra: https://mastra.ai
Learn Mastra in the world's first MCP-Based Course: https://mastra.ai/course
Principles of Building AI Agents (Book): https://mastra.ai/books/principles-of...
Patterns for Building AI Agents (New Book): https://mastra.ai/books/patterns-of-b...
MASTRA?
Mastra is an open-source TypeScript framework designed for building and shipping AI-powered applications and agents with minimal friction. It supports the full lifecycle of agent development—from prototype to production. You can integrate it with frontend and backend stacks (e.g., React, Next.js, Node) or run agents as standalone services. If you're a JavaScript or TypeScript developer looking to build an agentic or AI-powered product without starting from first principles, Mastra provides the scaffolding, tools, and integrations to accelerate that process.
📑 CHAPTERS
00:00 - Intro
02:17 - The history of coding agents: ChatGPT → Copilot → Aider → Claude Code
04:52 - What Claude Code got right — and where it became a spaceship
06:04 - Claude Code Drawbacks
09:39 - Claude Code Alternatives
11:38 - OpenCode's compaction problem and prompt cache busting
12:51 - Why LSP feedback mid-edit is a terrible idea
14:26 - OpenCode's architecture issues and security vulnerability
16:06 - TerminalBench and Terminus
18:13 - Mario's Two Theses
19:08 - Introducing pi — strip everything, build a minimal extensible core
20:01 - The system prompt
21:18 - What's not in pi — and what you build instead
22:40 - Extensions: custom tools, custom UI, hot reloading
24:00 - Community extensions
24:59 - Tree-structured sessions, cost tracking, HTML export
25:33 - TerminalBench results
25:54 - Open source under siege and human verification
It started in April 2025 when Peter Steinberger and Armin Ronacher (Flask, Sentry) dragged him into an overnight AI hackathon. Within weeks, Mario was hooked on Claude Code — until he wasn't. There was feature bloat, hidden context injection that changed daily, the infamous terminal flicker, and zero extensibility for power users.
He then surveyed the alternatives — Codex CLI, Amp, OpenCode... Eventually, he came across Terminus — an agent that gives the model nothing but a tmux session and raw keystrokes. If that's enough for the model to perform, what are all those extra features actually doing?
Mario's thesis: we're still in the "messing around and finding out" stage, and coding agents need to become more malleable so developers can experiment faster.
Pi is his answer: four tools (read, write, edit, bash), the shortest system prompt of any major agent, tree-structured sessions, full cost tracking, hot-reloading TypeScript extensions, and nothing injected behind your back. No MCP, no sub-agents, no plan mode — but all of it buildable in minutes through extensions.
The community has already shipped pi-annotate (visual frontend feedback), pi-messenger (a multi-agent chatroom), and someone even got Doom running. On TerminalBench, pi with Claude Opus 4.5 landed right behind Terminus — before it even had compaction.
🔗 LINKS & RESOURCES
pi coding agent: https://pi.dev
Mario Zechner: https://mariozechner.at
Peter Steinberger / OpenClaw: https://github.com/steipete
Armin Ronacher: https://lucumr.pocoo.org
Claude Code: https://docs.anthropic.com/en/docs/cl...
Aider: https://aider.chat
OpenCode: https://github.com/anthropics/opencode
Amp (Sourcegraph): https://sourcegraph.com/amp
TerminalBench: https://terminalbench.com
Ghostty: https://ghostty.org
Vouch: https://github.com/mitchellh/vouch
libGDX: https://libgdx.com
AI Engineer London is a community meetup for engineers and founders building with AI, covering everything from agent frameworks and RAG pipelines to LLMs in production. Each event features technical talks, live demos, and hands-on networking. This talk was recorded at AI Engineer London #10, hosted by Tessl, in collaboration with AI Engineer London.
AI ENGINEER LONDON
📅 Events: https://lu.ma/aiengineerlondon
💼 LinkedIn: / ai-engineer-london-meetup
📚 MASTRA RESOURCES
Mastra: https://mastra.ai
Learn Mastra in the world's first MCP-Based Course: https://mastra.ai/course
Principles of Building AI Agents (Book): https://mastra.ai/books/principles-of...
Patterns for Building AI Agents (New Book): https://mastra.ai/books/patterns-of-b...
MASTRA?
Mastra is an open-source TypeScript framework designed for building and shipping AI-powered applications and agents with minimal friction. It supports the full lifecycle of agent development—from prototype to production. You can integrate it with frontend and backend stacks (e.g., React, Next.js, Node) or run agents as standalone services. If you're a JavaScript or TypeScript developer looking to build an agentic or AI-powered product without starting from first principles, Mastra provides the scaffolding, tools, and integrations to accelerate that process.
📑 CHAPTERS
00:00 - Intro
02:17 - The history of coding agents: ChatGPT → Copilot → Aider → Claude Code
04:52 - What Claude Code got right — and where it became a spaceship
06:04 - Claude Code Drawbacks
09:39 - Claude Code Alternatives
11:38 - OpenCode's compaction problem and prompt cache busting
12:51 - Why LSP feedback mid-edit is a terrible idea
14:26 - OpenCode's architecture issues and security vulnerability
16:06 - TerminalBench and Terminus
18:13 - Mario's Two Theses
19:08 - Introducing pi — strip everything, build a minimal extensible core
20:01 - The system prompt
21:18 - What's not in pi — and what you build instead
22:40 - Extensions: custom tools, custom UI, hot reloading
24:00 - Community extensions
24:59 - Tree-structured sessions, cost tracking, HTML export
25:33 - TerminalBench results
25:54 - Open source under siege and human verification
2026/w15/i-hated-every-coding-agent-so-i-built-my-own-mario-zechner-p