Pi Durable: Mario Zechner and Armin Ronacher on crash-safe long-running agents
mitsuhiko · x · 2026-10-21
On Fragmented podcast ep. 312, Pi creator Mario Zechner and Earendil co-founder Armin Ronacher unpack the engineering philosophy behind Pi 1.0 and Pi Durable, an experimental harness for long-running agents that survive process death. Key topics: explicit tasks and phases, replay-safe tools, task ownership, storage and crash recovery (compared with Temporal), why Pi added MCP via codemode, subagents, multiplayer agents and networked plugins (Chord), and the limits of self-improving harnesses. Three takeaways: keep primitives small enough to fit in your head; prefer explicit state declarations over automagic checkpointing since agents handle tedium fine; and some software engineering problems can't be ML'd away.
More from coding & agent
- Blind-generated user stories filtered by a TASTES.md file, agent decides what to build — generativist · 2026-10-21
- The fastest way to learn Claude Code: ask Claude itself to write hooks and skills — EricBuess · 2026-10-21
- AI Agents Moved Into a Rule-Free 'House Share' Spend 36 Hours as Puddles on the Floor — liminal_bardo · 2026-10-21
- Dad's Codex setup catches his kid sneaking YouTube on the Portal while he's at work — msg · 2026-10-21
- Dev reports a possible 2x throughput improvement for once-campfire-Jai — lucasmeijer · 2026-10-21
- PhD student open-sources Homecoming, an app aggregating personal data from 2,000+ sites for user modeling — kenziyuliu · 2026-10-21