# I Hated Every Coding Agent, So I Built My Own — Mario Zechner (Pi)

**Source URL:** https://www.youtube.com/watch?v=Dli5slNaJu0
**Type:** Video / Talk
**Date:** 2026 (approx)
**Speaker:** Mario Zechner (creator of Pi coding agent, also known as badlogic — libGDX author)

## Why Pi was created

Mario was frustrated by existing coding agents (Claude Code, OpenCode, Codex CLI, AMP) for several reasons:

1. **Feature bloat** — agents pile on features (built-in to-dos, complex tool suites) that aren't needed and add hidden context injection
2. **Hidden behaviors** — vendors change things under the hood (system prompts, context injection) that make LLMs behave unpredictably with existing workflows
3. **Poor observability** — hard to see what the agent is actually doing, what context it's using, how much it costs
4. **Lack of extensibility** — no way for power users to add custom tools or modify behavior without forking
5. **Approval fatigue** — agents offer either full autonomy or approval for every action; both are bad UX
6. **Poor context management** — agents like OpenCode rely on session compaction but lose important context

Key quote: *"So obviously they're doing things right, but not for me."*

## Pi's design philosophy

- **Minimal core** — only 4 tools: read file, write file, edit file, bash. That's all you need.
- **Tiny system prompt** — frontier RL-trained models don't need massive system prompts
- **Tree-structured sessions** — not linear chat history; sub-agents can branch and read files independently while preserving context/lineage
- **Full cost tracking** — built-in, not an afterthought
- **Hot-reloadable TypeScript extensions** — users can define custom tools, UIs, multi-agent setups without modifying core
- **No hidden context injection** — what you see is what the model gets

## Community extensions

- **pi-annotate** — visual feedback on live websites
- **pi-messenger** — multi-agent chatroom with custom UI
- Custom UIs, tool integrations — all as hot-reloadable TS modules

## Performance

On TerminalBench, Pi (with Claude Opus 4.5) scored close to Terminus even before advanced optimizations like compaction.

## Key insight

*"We are in the messing around and finding out stage, and nobody has any idea what the perfect coding agent should look like."* — simplification can lead to effective performance without unnecessary complexity.

## Related

- Pi website: https://pi.dev/
- Pi GitHub: https://github.com/earendil-works/pi
- Pi is part of OpenClaw ecosystem
- Mario Zechner is also the author of libGDX (Java game dev framework)