Pi agent: the case for simpler coding-agent harnesses

In a David Andre podcast, Armin Ronacher argues that capable coding models often need a small, composable tool surface rather than a large proprietary harness. Pi leans on Bash and Unix-style pipelines, while the harder next problems are durable sessions, portable state, data access, and interfaces that people can inspect and use. Why it matters: More agent features do not automatically mean better work. A reliable environment, clear approval boundaries, and tools that compose may matter more than a crowded interface—especially when teams need to understand and own what an agent did.

Try this: Take one recurring engineering task and run it with your current agent setup and a minimal shell-first workflow. Compare context use, reproducibility, reviewability, and the quality of the final result—not just speed.