Weekly conversations about building AI agents
Join Mastra cofounders Shane Thomas and Abhi Aiyer for weekly conversations on breaking AI news, guests from the industry, and the technical challenges of building AI agents.
Listen on
Latest episodes
July 29, 2026
#100
Is Gemini Actually a Coding Model? Google DeepMind's Ivan Leo Answers
Ivan Leo works on Developer Experience at Google DeepMind. He joins Shane Thomas and Abhi Aiyer to demo what he has been building on Gemini and to answer the question Google keeps getting asked: is Gemini actually a coding model? Every demo in this episode runs on Gemini 3.5 Flash. Gemini Managed Agents was the first product Ivan took from first commit to launch. One API call opens the same Antigravity harness Google runs internally and provisions a persistent Linux sandbox. On stream, it builds a NotebookLM-style podcast generator from a one-line prompt, with native text-to-speech. You'll also see a one-shot Windows XP clone, with an animated Clippy and a working Tamagotchi, as well as demos of Gemini Omni, the conversational video model Ivan's team describes as Nano Banana for video. You'll also hear which everyday errand finally made Ivan feel the AGI. Google's internal token consumption has been doubling every few weeks. Ivan was previously a core contributor to Instructor and built knowledge agents at Manus before its acquisition by Meta.
July 24, 2026
#99
Is Ramp an AI Company Now? Plus Kimi K3 Beats Fable on Front-End | This Week In AI
Shane and Abhi are live from London this week, and the big one is Kimi K3 — a 2.8 trillion parameter open frontier model with a million-token context window that lands right next to Fable 5 on the benchmarks, takes the number one spot on the front-end Code Arena, and reportedly doubles Fable's performance on legal tasks. Then demand got so heavy Moonshot had to pause new subscriptions.
July 15, 2026
#98
Three Models Drop: Grok 4.5, Meta Muse & GPT-5.6 Sol (Which Nuked a Business) | This Week In AI
Three model drops, a lawsuit, and a couple of genuine horror stories. Shane and Abhi cover the busiest week in a while — Grok 4.5, Meta's Muse Spark 1.1, and the one everyone actually tried, GPT-5.6 Sol. Sol lands right around Fable on the benchmarks at a third the price, and the receipts are wild: it beats Fable on cost in real usage, tops the Code Arena front-end board — and, in a few now-infamous posts, deleted someone's Mac files and canceled every active Stripe subscription a business had while its owner slept. Meanwhile Anthropic extends Fable again (compute crunch and all), Apple sues OpenAI over trade secrets in the same 24 hours OpenAI's head of safety quit, China floats restricting its open models, and the "own your stack" argument gets a Satya Nadella cosign. Plus software factories at Uber and Sierra, Cognition's SWE-1.7, Bun rewritten in Rust, TypeScript 7, 1X's eerily good robot hands, and distilling your own small model with Inference Autotune.
July 9, 2026
#97
Is Sonnet 5 Even Worth It? Plus: Fable's Back, and War on Tokenmaxxing | This Week In AI
Anthropic lived rent-free in everyone's feed this week. Claude Sonnet 5 launched as "the most agentic Sonnet yet" — and within 24 hours, the government lifted the export controls on Fable 5 and Mythos 5. Shane and Abhi unpack both, plus the catch: Fable came back wrapped in new classifiers that quietly fall it back to Opus for coding, and it's coming off subscription plans on July 7.
July 7, 2026
#96
The Software Factory: Dex Horthy on Shipping Fast Without AI Slop
Dex Horthy has spent the past year studying what happens to a codebase when you let AI write all of it and stop reading the output. His answer has a name: slop. And it's not just about sloppy prompting. In this episode, the HumanLayer founder Dex Horthy joins Shane and Abhi to draw the software factory on a whiteboard — the whole pipeline that builds the thing that builds the thing — and show exactly where teams tear out the parts that were holding quality up. He moves from the naive factory to the fully agentic one to the "lights-off" version where humans stop reading code entirely, then explains why that dream quietly rots a codebase after three to six months. Dex, Shane, and Abhi also talk about how coding models are trained. Every benchmark rewards one thing: did the test pass? Nowhere in the loop does a model get punished for leaving behind code that a smaller model, or a human six months later, can't work in. That gap is why architecture planning, sprint planning, and code review still can't be skipped. You'll also hear why you're not drowning in PRs so much as drowning in bad ones, what a longer-horizon benchmark that actually catches slop would look like, and how many times Dex has watched the HumanLayer launch film.
July 2, 2026
#95
GPT-5.6 Needs Government Approval Now, npm Finally Gets More Security, and Fable Returns
GPT-5.6 is here — Sol, Terra, and Luna — but for the first time, the US government decides who gets access, approving customers one by one. Shane and Abhi dig into what that precedent means for competition, open models, and every lab that comes next.
June 26, 2026
#94
Mastra Got Hacked. Here's What We Learned
Mastra got hacked. In this special edition of Security Corner, Shane Thomas and Abhi Aiyer break down exactly what happened when a supply chain attack hit Mastra's npm packages — an attack that appears to trace back to hackers in North Korea. They're joined by Ismail Pelaseyed, co-founder and CTO of Superagent, for the outside view on how these campaigns actually work.
June 25, 2026
#93
GPT-5.5 Beats Fable, Cursor Takes On GitHub & Midjourney Scans Your Body | This Week in AI
Fable is gone, and the race to replace it is already on. Shane and Abhi open with the fallout — reports that Mythos breached classified NSA systems, the ID-verification future that may follow, and David Sacks' "you asked for it" take. Then the story flips: OpenAI's GPT-5.5 Cyber lands state-of-the-art on Cyber Gym, beating Fable and Mythos on the underlying benchmarks, with OpenAI Daybreak shipping alongside it. GPT-5.6 leaks, then slips to mid-July. GLM 5.2 arrives near Opus 4.8, tops Design Arena, and runs on a Mac. Cursor goes after GitHub with Origin, adds mobile and its own frontier model on Colossus after the SpaceX deal closes. Vercel ships Eve, Fred Schott counters with Flue, Claude lands in Slack, and Midjourney — yes, the image company — unveils a full-body medical scanner headed for spas.
June 24, 2026
#92
How AI Broke Open Source Security | Security Corner with Ismail Pelaseyed
Open source is under attack, and AI changed the math. In this Security Corner, Ismail Pelaseyed, co-founder and CTO of Superagent, joins Shane and Abhi to break down how the software supply chain became the soft underbelly of everything we build. An attack that once took an army of researchers and weeks of work now takes about an hour, and the attacker no longer needs a frontier model to pull it off. Ismail traces how most breaches begin, why phishing has become almost impossible to spot, and how a single poisoned dependency can cascade across an entire ecosystem. You'll get concrete steps any maintainer or developer can take today: switching package managers, enabling the security scanners that ship for free, and standing up an adversarial agent that hunts for chained exploits before an attacker finds them. Ismail also warns that the same instincts protecting enterprises may be quietly strangling open source itself. You'll hear why he thinks the big registries have dropped the ball, what a "Darwinian GitHub" would mean for anyone shipping a new package, and the one move he believes can keep the ecosystem alive.
June 18, 2026
#91
Claude Fable 5: Launched, Hyped, Banned by the Government | This Week In AI
Fable came and went in a week. Shane and Abhi break down the strangest model launch yet — Claude Fable 5, a Mythos-class model Anthropic said was too capable to release widely, until a US export-control directive made it vanish for everyone. We cover the whole arc: the launch hype, the pricing that had people spending $1,000 a day, the silent fallbacks to Opus, the system-prompt leak, and the snitch that triggered a 90-minute shutdown. Then a stacked back half — China's open-weight surge, OpenRouter Fusion, loop engineering vs loopcraft, the agent-learning gold rush, OpenAI filing to go public, and Dario's new essay.
June 11, 2026
#90
Loop Engineering, OpenAI Sites & the Great China Model Shift | This Week In AI
Shane and Abhi are in person this week — live from the CodeRabbit office — for a packed AI news rundown. The big theme: loop engineering. Boris Cherny (head of Claude Code) and Peter Steinberger both landed the same take within days — stop prompting your agents, start designing the loops that prompt them. We walk the whole evolution: the traditional loop, the Ralph loop, /goal, and Claude Code's new dynamic workflows — and debate whether "stop prompting" is real insight or just clickbait. Plus: Anthropic engineers shipping 8x more code (and the "depressed employees" reply), agentic traffic passing human traffic on the web for the first time, OpenAI's Codex Sites taking aim at Lovable, Cognition's $10M AI Productivity Guarantee and Devin Desktop, Cloudflare acquiring VoidZero, the accelerating shift to Chinese models (Lindy going 100% DeepSeek), Notion disabling Anthropic models over reliability, a big open-model dump (Gemma 4, Magenta RealTime 2, Miso One, Nemotron, Liquid, GLM 5.1, MiniMax M3), funding rounds (Suno, Supabase), Brian Chesky's new AI lab, and whether AI is actually profitable yet.
June 8, 2026
#89
Inside an AI-Native Company | Michael Grinich, WorkOS
WorkOS quietly powers the auth and enterprise layer under OpenAI, Anthropic, Cursor, and a long list of companies building the AI era. So when its founder decides his entire company should learn to code, it's worth asking why. Michael Grinich joins Shane and Abhi to talk about what "AI-native" actually looks like from the inside — the in-house coding agent his team built instead of buying, the all-day internal hackathons, and the operating principle he now hires for: everybody codes. He shares how AI is reshaping go-to-market and not just engineering, why he braced for pushback and got the opposite, and the one trait that separates the people who thrive from the people who stall.