The Frontier

Mario Zechner argues current models lack sufficient RLHF data on software architecture and design, making them ineffective at structuring solutions.

Zechner uses agents on modular, well-architected code where boundaries are clear, but reserves final oversight for mission-critical and security-related components.

Zechner built Pi, a minimalist coding agent harness based on a small, extensible core that users can modify themselves to fit workflows, opposing heavy feature-driven designs.

Zechner avoids MCP integrations in Pi, citing issues with server implementations wasting context tokens on tool definitions and preferring direct CLI use.

Zechner's workflow for bug fixes includes using Pi with an issue prompt template to fetch, label, and analyze GitHub issues, verifying the analysis before implementing.

Zechner manually reviews agent-generated code to combat unnecessary abstraction and complexity, using a custom Pi extension to provide inline feedback.

Zechner's agents.md file defines coding style and rules, but notes models often ignore it, relying more on deterministic linting and type-checking for enforcement.

Zechner says agents can massively degrade a codebase faster than human teams, requiring ruthless refactoring, but believes they can also assist in that cleanup.

Zechner uses GPT-5.5 as his daily driver for code but switches to Claude for prose, and dabbles with open-weight models like Kimi 2.6 and DeepSeek.

Zechner avoids automatic worktree creation in Pi, citing distrust of models handling complex git operations and relying on modular code to prevent file conflicts.

Zechner refactors large codebases by first using the agent to explore and summarize relevant files, then carrying that summary into a separate implementation branch within the session.

Zechner built a robot with a Pi brain over 12 hours, using voice-to-text and agent-generated frontend code, then refactored the messy result by modularizing tool implementations.

Zechner advocates adversarial agent roles to push back on user ideas and prevent sloppy code, referencing Matt Shumer's 'roast me' skill as an example.