311 - Self learning harnesses are here

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 NotesPi 1.0, Earendil and Pi DurableMario 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 recoveryPi 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 pluginsPi 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 limitsEpisode 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 developersEarendil - 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 touchPlease 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 liketo hear more on.Contact usNewsletterYoutubeWebsiteCo-hosts:Kaushik GopalIury Souza

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 everything — down to the agent loop itself — is a plugin you can hot-swap at runtime, and Prime Agent, an RLM-based harness that treats sub-agents and context as function calls in a live REPL. We break down what actually makes each one different.Full shownotes: https://fragmentedpodcast.com/episodes/311

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 coding actually happens. Mitchell was an AI skeptic. We walk through his six-step adoption framework and the workflows he uses day to day — warm-start research, Hail Mary prompts across twenty GitHub issues, and knowing when to let the agent slam dunk it. Full shownotes at fragmentedpodcast.com. Show Notes HashiCorp Vagrant Terraform IBM acquires Hashicorp Ghostty Ghostty - Mitchell's fast, native terminal built for platform integration across Mac and Linux Terminal shell SSH - secure shell PTY - pseudoterminals Terminal Multiplexers tmux - most popular open source one XTGETTCAP by xterm libghostty - the cross-platform terminal emulation library that powers Ghostty's core xterm-js - powers terminal for apps like VSCode and the cloud Jedi Term - Intellij's embedded terminal Ghostty is now a non-profit cmux - native macOS terminal multiplexer built on libghostty — a fork Mitchell champions Free Software Definition - the 4 essential freedoms The freedom to run the program as you wish, for any purpose. The freedom to study how the program works, and change it to make it do what you wish. The freedom to redistribute copies so you can help others. The freedom to distribute copies of your modified versions to others. Mitchell's tweet on unsolicited PRs and transfer of ownership The AI Adoption Journey My AI Adoption Journey - Mitchell's blog post outlining his five-step framework Step 1: Drop the Chatbot Episode 301 - AI Coding ladder - Different stages of AI adoption Step 2: Reproduce Your Own Work Step 3: End-of-Day Agents OpenAI Deep Research - kick off research tasks for a "warm start" the next morning Spine AI research - deep research tool for longer, hour-long analysis tasks Step 4: Outsource the Slam Dunks Claude status hooks - warcraft peons Conductor Step 5: Engineer the Harness Episode 307 - Harness Engineering - Fragmented's deep dive on harness engineering, heavily inspired by Mitchell's post Step 6: Always have an Agent running Peter Steinberger Codex plugin for Claude Code Get in touch We'd love to hear from you. Email is the best way to reach us or you can check our contact page for other ways. 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 [!fyi] We transitioned from Android development to AI starting withEp. #300. Listen to that episode for the full story behind our new direction.

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 four parallel Gradle builds, and review queues back up. Kaushik and Iury trace the full arc — from local multitasking to cloud-hosted async work to fully autonomous agents that fire on repo events and put PRs in your inbox. Show Notes Andrej Karpathy on agents and token throughput - NoPriors podcast — maximize agent runtime, not token burn Cursor Agent Mode - Multiagent interface - introduced the multi-agent board as a new paradigm for local parallel agents Google Antigravity - Agent Manager interface Claude Code Agent Teams - spawn sub-agents from a main orchestrator, with tmux pane integration Git worktrees - /reddit Remote Background Agents in the cloud Google Jules - hosted GitHub-connected agent, proposes a plan, edits code, runs tests, opens a PR Cursor Cloud Agents - remote agents that clone your repo in the cloud and work in parallel OpenAI Codex - cloud software engineering agent for parallel tasks Claude Code on the web - cloud-hosted Claude Code sessions decoupled from your local machine Building trust Episode 307 - Harness Engineering - the earlier episode on shaping agent environments — and why this ceiling exists Get in touch We'd love to hear from you. Email is the best way to reach us or you can check our contact page for other ways. 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 [!fyi] We transitioned from Android development to AI starting withEp. #300. Listen to that episode for the full story behind our new direction.

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 actually works — and what it has in common with the text models you already understand. Full shownotes at fragmentedpodcast.com. Show Notes Episode 303 - How LLMs work in 20 minutes - text generation VAE -Variational Autoencoder RGB Color model - wikipedia Word2Vec technique - wikipedia Efficient Estimation of Word Representation - original Word2Vec paper by Mikolov et al. High-Resolution Image Synthesis with Latent Diffusion Models - Rombach et al. (2022) — the paper behind Stable Diffusion Image Training data LAION-5B - 5 billion image-text pairs scraped from the web, used to train many image generation models WebLI - Google's internal image-text dataset Michelangelo Get in touch We'd love to hear from you. Email is the best way to reach us or you can check our contact page for other ways. 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 [!fyi] We transitioned from Android development to AI starting withEp. #300. Listen to that episode for the full story behind our new direction.

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 when teams build custom harnesses from scratch. Full shownotes at fragmentedpodcast.com. Show Notes Why it matters Harness Engineering - OpenAI's post on building their Codex codebase (~1M lines of code, 1,500 PRs merged, zero manually written) Shaping the harness The Feed's Lost and Found - Iury's newsletter consolidating harness engineering themes Agent legibility Closed feedback loops Persistent memory Entropy control Blast radius controls Building the harness Minions: Stripe's one-shot, end-to-end coding agents - Stripe forked Goose to build custom agents for their codebase Goose - open-source coding agent from Block Superpowers by Jesse Vincent - skills that enforce a proper software engineering process Open Code - open-source coding agent you can fork and customize Other resources Agent Harness Glossary - Latent Patterns Towards self-driving codebases - Cursor Agentic Workflows - GitHub Next Future of Software Development - ThoughtWorks Get in touch We'd love to hear from you. Email is the best way to reach us or you can check our contact page for other ways. 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 [!fyi] We transitioned from Android development to AI starting withEp. #300. Listen to that episode for the full story behind our new direction.

AGENTS.md is becoming the common language for AI coding tools, but keeping repo rules, personal rules, and tool-specific files in sync is still messy. In this episode, Kaushik and Iury break down the sync problem, compare their own setups, and unpack what the latest AGENTS.md research actually says. Full shownotes at fragmentedpodcast.com. Show Notes The sync problem AGENTS.md - Official spec Custom instructions with AGENTS.md - Open AI Keep your AGENTS.md in sync - Kaushik's post Rulesync - What Iury uses Tweet by Ryan Carson and Claude frustrations Other links Evaluating AGENTS.md: Are Repository-Level Context Files Helpful for Coding Agents? Harness engineering - Check the section about using AGENTS.md as a table of contents OpenCode Get in touch We'd love to hear from you. Email is the best way to reach us or you can check our contact page for other ways. 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 [!fyi] We transitioned from Android development to AI starting withEp. #300. Listen to that episode for the full story behind our new direction.

Subagents are becoming a core primitive for serious AI-assisted development. In this episode, Kaushik and Iury disambiguate "agent" terminology, unpack plan mode vs subagents, and explain how parallel, scoped workers improve research quality without polluting the main thread.Full shownotes at fragmentedpodcast.com.Show NotesResources & DocumentationOfficial DocumentationAgents, Modes, Subagents: official harness docsClaude Code SubagentsGemini CLI SubagentsOpencode SubagentsCursor SubagentsAntigravity Agent ModesAOE ScoutingResearch Papers & ArticlesIntroducing Claude Opus 4.5Deep-Research Agents PaperPost: GPT-5 System Card by AlexXuSelf-Driving Codebases Blog -multi-agent systems making 1,000 commits/hourGet in touchWe'd love to hear from you. Email is thebest way to reach us or you can check our contact page for otherways.We want to hear all the feedback: what's working, what's not, topics you'd liketo hear more on.Contact usNewsletterYoutubeWebsiteCo-hosts:Kaushik GopalIury Souza[!fyi] We transitioned from Android development to AI starting withEp. #300. Listen to that episode for the full story behindour new direction.

Agent Skills look simple, but they are one of the most powerful building blocksin modern AI coding workflows. In this episode, Kaushik and Iury break down whento use skills, how progressive disclosure works, and how skills compare withcommands, instructions, and MCPs.Full shownotes at fragmentedpodcast.com.Show NotesMain ReferencesProgressive Disclosure -how skills are loaded into contextAgent Skills Open SpecificationAAIF (Agentic AI Foundation) -Linux Foundation initiative for AI interoperabilityNeedle in a Haystack Problem - original"Lost in the Middle" paperAgent-Invokable vs User-Invokable -merging skills and commands in Claude CodeCreating SkillsSkill Creator -Anthropic's skill for creating new agent skillsClaude Code frontmatter referencesee model: * & context: forkUsing other SkillsAnthropic Skills GitHub Repository -official collection of Claude skills and examplesClawdhub - Clawdbot's skill hub. All versions arearchived hereSKILLS.sh - Vercel's skills hubWarnings before installing random skills[!warning] Don't install from hubs blindly.Inspect the repo code before adding anything to your agent.Prompt Injection Attacks -OWASP guide to LLM prompt injection vulnerabilitiesOpenClaw OpenClaw Security Analysis -analysis of prompt injection risks in open agent frameworksMalware found in a top-downloaded Clawhub skill -incident report threadAdditional resourcesFew-Shot Prompting -improving outputs with examples.agents/skills - proposalto standardize the skills folder pathVercel: AGENTS.md vs Skills -comparison of agent instruction methodsGet in touchWe'd love to hear from you. Email is thebest way to reach us or you can check our contact page for otherways.We want to hear all the feedback: what's working, what's not, topics you'd liketo hear more on.Contact usNewsletterYoutubeWebsiteCo-hosts:Kaushik GopalIury Souza[!fyi] We transitioned from Android development to AI starting withEp. #300. Listen to that episode for the full story behindour new direction.

Ever get asked "how do LLMs work?" at a party and freeze? We walk through the full pipeline: tokenization, embeddings, inference — so you understand it well enough to explain it. Walk away with a mental model that you can use for your next dinner party._Full shownotes at fragmentedpodcast.com.Show NotesWords -> Tokens:OpenAI Tokenizer visualizer -Visualize how text becomes tokensTokens -> Embeddings:RGB Color model - wikipediaWord2Vec technique - wikipediaEfficient Estimation of Word Representation -original Word2Vec paper by Mikolov et al.Embeddings -> Inference:Word embeddingTemperature, Top-k, Top-p sampingGet in touchWe'd love to hear from you. Email is thebest way to reach us or you can check our contact page for otherways.We want to hear all the feedback: what's working, what's not, topics you'd liketo hear more on. We want to make the show better for you so let us know!Contact usNewsletterYoutubeWebsiteCo-hosts:Kaushik GopalIury Souza[!fyi] We transitioned from Android development to AI starting withEp. #300. Listen to that episode for the full story behindour new direction.