Skip to content
PodSubs
ShowsPeopleGuestsJobsNewsEvents
More
BlogPodCred
Menu
ShowsPeopleGuestsJobsNewsEventsBlogPodCredNetworksSign in
Join nowSign in
  • Shows
  • People
  • Jobs
  • News
  • More
    • Guests
    • Events
    • Blog
    • PodCred
    • Networks
    Join nowSign in

Fragmented - AI Developer Podcast

Fragmented - AI Developer Podcast artwork

312 - Understanding Pi Durable with Mario & Armin of Earendil

Oct 20, 20261 hr 1 minAudio

What does it take to keep an agent working after its process dies? Pi creator Mario Zechner and Earendil co-founder Armin Ronacher join KG to explain Pi Durable: explicit tasks, crash recovery, and the primitives for multiplayer agents. We also dig into why Pi added MCP and codemode, and why a harness that can rewrite itself still needs a clear goal. Full shownotes at fragmentedpodcast.com. Show Notes Pi 1.0, Earendil and Pi Durable Mario Zechner - Pi’s creator and our guest. Armin Ronacher - Earendil co-founder and our guest. Pi 1.0 - The stable release and the design choices behind it. Pi Durable - Mario’s introduction to the experimental harness for long-running agents. Codemode and MCP: why Pi added both “You Said No MCP!” - Why codemode brought MCP and non-LLM models into Pi. Blender MCP - The Blender integration Armin uses as an MCP example. Durable workflows and crash recovery Pi Durable documentation - Tasks, explicit phases, replay-safe tools, ownership, and storage. Temporal - The workflow engine compared with Pi Durable’s explicit task model. Absurd Workflows: Durable Execution With Just Postgres - Armin’s earlier workflow design, and the trade-offs discussed here. Subagents, multiplayer and networked plugins Pi Durable: abort and subagents - How task ownership connects parent work and child conversations. Chord - Networked objects, versioned contracts, and hot reload across clients. Self-improving harnesses and their limits Episode 311 - Self learning harnesses are here - The previous discussion that prompted KG’s questions. Prime Agent - The harness discussed while questioning goals, evaluation, and unconstrained rewrites. Terminal-Bench - The benchmark used to explain narrow optimization and benchmark gaming. Earendil and advice for developers Earendil - The team’s broader goal: AI tools that work beyond software engineering. Bryan Cantrill’s The Observation Deck - The writing Armin recommends to software engineers. Reflecting on Life - Armin’s broader advice on work, commitment, and personal growth. Get in touch Please check our contact page for the best ways to reach out to us. We want to hear all the feedback: what's working, what's not, topics you'd like to hear more on. Contact us Newsletter Youtube Website Co-hosts: Kaushik Gopal Iury Souza

Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.

Hosts and Guests

Nobody is credited on this episode yet.

Credit the people on this episode

Be the first to rate this episode

More episodes

All episodes
  1. 311 - Self learning harnesses are here

    Sep 6, 202613 minAudio

    Claude Code and Codex have been the mainstay agent harnesses, with OpenCode and Pi as the popular open-source alternatives. But two new entries are pushing what a harness even is: DeepSeek's Cordis kernel, where...

    Credit the people on this episode

  2. 310 - Mitchell Hashimoto on Ghostty & His Agentic Coding Workflow

    Apr 14, 20261 hrAudio

    Mitchell Hashimoto co-founded HashiCorp, built some of the most impressive DevOps tools like Vagrant and Terraform, sold the company to IBM - and then built a terminal. Ghostty is now where a huge chunk of agentic...

    Credit the people on this episode

  3. 309 - Background Agents

    Apr 1, 202626 minAudio

    Andrej Karpathy says the goal is to maximize how long an agent runs without your intervention. But there's a false summit most teams hit first: individual speed goes up while system speed stalls, your laptop roars under...

    Credit the people on this episode

  4. 308 - How Image Diffusion Models Work - the 20 minute explainer

    Mar 24, 202625 minAudio

    You already know how LLMs work from our popular 20-minute explainer. Now we take it to images. What does Michelangelo have to do with stable diffusion? More than you'd think. Walk away knowing how image generation...

    Credit the people on this episode

  5. 307 - Harness Engineering - the hard part of AI coding

    Mar 17, 202630 minAudio

    The hard part of AI coding isn't generating code - it's controlling quality, safety, and drift. Kaushik and Iury break down harness engineering: the five pillars for shaping an agent's environment and what it looks like...

    Credit the people on this episode