Building Pi, and what makes self-modifying software so fascinating

Mario Zechner and Armin Ronacher discuss Pi and self-modifying coding-agent harnesses. They describe engineering practices for when agents make code cheap but judgment, review, and responsibility still do not scale.

Key Points Covered

• Agent adoption takes immersion: Ronacher's interviews with roughly 30 engineering teams suggest agents usually click after a couple of weeks, often during holidays or downtime. He reports that wider adoption also brings bigger PRs, more review load, and early signs of "vibe slop" in production code 15:14–18:03(opens in a new tab).
• Automation bias hides complexity until humans feel pain: Zechner and Ronacher contrast agents with humans. They say agents can alternate between good and terrible output and do not learn from codebase pain. They also argue agents make it easier to approve features instead of using senior engineering judgment to reject them 19:58–23:34(opens in a new tab).
• Non-engineer code still needs guardrails: The discussion treats PM-, sales-, and marketing-generated code as empowering for prototypes. It warns that "prompt requests" or rough implementations should clarify intent rather than bypass engineering ownership 24:21–27:20(opens in a new tab).
• Pi came from wanting a stable, controllable harness: Zechner says Claude Code's hidden context changes and shifting prompts broke his workflows. He says OpenCode's context pruning and eager LSP diagnostics pushed him toward a simpler tool he could understand and adapt without forking 32:06–37:26(opens in a new tab).
• Pi's product surface is malleability: Its core is a provider abstraction, agent loop, TUI, and minimal tools such as read, write, edit, and bash. Hook points let users ask Pi to add MCP support, plan modes, UI changes, or custom tools itself 37:26–40:02(opens in a new tab).
• Specialized harnesses keep humans and agents in the loop: Ronacher describes making Pi build game-specific debugging, screenshots, simulation dumps, and branching workflows. These let both the agent and human validate progress in the least confusing way for that project 40:53–42:49(opens in a new tab).
• Self-modifying software points beyond terminals: Zechner argues the broader pattern is software that modifies itself around a user's task. He presents Pi as a coding-agent example and says future web interfaces offer more room than line-based TUIs 45:37–47:17(opens in a new tab).
• Open source needs bottlenecks against agent spam: OpenClaw's use of Pi exposed Zechner's repo to autonomous issues and PRs. He then auto-closed PRs from unrecognized accounts and required a short, human-voiced issue before allowlisting contributors 48:06–53:34(opens in a new tab).
• Complexity is still the agent's enemy: The speakers argue that agents can only operate on the context they find and fit. As agents add interconnected code, they make future tasks harder. The speakers say model defaults tend toward internet-average patterns rather than the best-engineered projects 1:00:15–1:02:20(opens in a new tab).
• "Slow down" means using agents where review scales: Zechner warns that more generated code means more errors to review, not less responsibility. He argues teams should automate annoying work, preserve time for product judgment, and keep important code under human refactoring discipline 1:10:38–1:16:19(opens in a new tab).
• MCP and CLIs solve different problems: Ronacher and Zechner see MCP as useful for auth and enterprise integration, but criticize context-heavy, non-composable tool APIs. They say CLIs and code execution let agents combine data through pipes or scripts while exposing only the relevant result 1:16:19–1:24:39(opens in a new tab).

Full video: https://www.youtube.com/watch?v=n5f51gtuGHE(opens in a new tab)