Fragmented - AI Developer Podcast

305 - Subagents explained - What they are, when (not) to spawn them

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.

304 - Agent Skills - when to use them and why they matter

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

303 - How LLMs Work - the 20 minute explainer

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.

302 - MCPs Explained - what they are and when to use them

MCPs are everywhere, but are they worth the token cost? We break down what Model Context Protocol actually is, how it differs from just using CLIs, the tradeoffs you should know about, and when MCPs actually make sense for your workflow.

301 - The AI coding ladder

Most folks reference "AI coding" like it's one thing. It's really not. In this foundational episode Kaushik & Iury walk through (at least) four paradigms — from super autocomplete to agent orchestration — each with different workflows, expectations, and mental models.What do most developers follow today? Where is the frontier? What's coming in the future?Listen to the episode and find out!Full shownotes at fragmentedpodcast.com.Show NotesGen 1: Super autocompleteIntellisense - regular autocompleteGithub CopilotCursor TabGen 2: Chat Oriented ProgrammingCursor IDEFirebenderGen 3: AgentNvidia's definition of an AgentReAct PromptingChain of Thought was a prompting hackDeepSeekDeepSeek - R1 paperTUI tools (or Harnesses):Claude CodeOpen CodeCodex CliGemini CliIDE style toolsCursor AgentCopilot (MS)Junie - IntellijAntigravity - GoogleHeadless Tools:Jules - GoogleClaude Code on the WebCodex WebGen 4: Agent OrchestrationGit worktreesTipsIury: Transfer between agents using your own compact commandKG: Ask the agent to clarify your promptConfirm if my requirements are clear. If you have follow up questions, ask mefirst and clarify before executing anything.Contact usNewsletterWebsiteContact usYoutubeCo-hosts:Kaushik GopalIury Souza