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
September 8, 2026
#112
Multiplayer Coding Agents in the Cloud - Charlie Holtz, Conductor
"Code is almost like the sawdust of the conversation." Charlie Holtz, co-founder and CEO of Conductor (YC S24), returns to Agents Hour to show how far coding agents have come in just a few months. Conductor started as a Mac app for running a handful of Claude Code agents in parallel. Now it's fully cloud-native and multiplayer! According to Charlie, once Fable-class models could run for hours or days, keeping agents chained to your laptop stopped making sense. Charlie gives a live demo of what he thinks everyone in cloud is overlooking: multiplayer. See every teammate's agents running in real time, jump into a workspace and steer an agent together, hand off a whole session with one keystroke, and watch agents spin up their own workspaces through Conductor's API. Along the way, Shane and Abhi dig into the ideas underneath it: code as a disposable byproduct of the prompt, the Figma-style collaboration unlock, why today's models aren't trained for multi-person conversations, and why Conductor built a native iOS app instead of just living in Slack.
September 3, 2026
#111
OpenAI Cuts Off Cursor, Nvidia Buys Hugging Face, Ox Alpha is GLM | This Week In AI
Someone became worth $2.5 billion in a week, and that's just the warm-up. Shane and Abhi open on the drama of the week: OpenAI cutting off Cursor after its acquisition by SpaceX, telling the team direct access to its models ends November 12th over "trust." Cursor's Michael Truell waves it off as 5% of traffic, Anthropic jumps in to say it would never do such a thing, and Amjad reminds everyone about Windsurf. It's a Game of Thrones, and meanwhile China is making moves. Ox-alpha, last week's stealth-model mystery, got unmasked: it's GLM-5.3, now multimodal, MIT-licensed, open weight, and reportedly training on Chinese chips. Then Instinct, a personal agent nobody had heard of, launches and hits a $2.5B valuation in a week, real-time "infinite video" turns Twitch into interdimensional cable, and Dwarkesh's viral "secret AI civilizations" thread gets a reality check from the people who actually read the report. Plus Claudeforce puts Claude natively inside Salesforce (with a Matthew McConaughey ad), NVIDIA agrees to buy Hugging Face for $12.9B, Hugging Face unveils a $399 Microduck robot, and quick hits on Crof, ElevenLabs CLI v1, and DuckLabs joining AWS.
August 28, 2026
#110
Why Agents Are Actually Workflows - Tony Kovanen, Mastra's Founding Engineer
What if agents and workflows aren't two different things — but the same thing? Abhi sits down with Tony Kovanen, Mastra's founding engineer, live in London, to explore the ideas behind Mastra's design. Tony co-created Next.js and was a core team member of Socket.IO before joining Mastra as employee number one, and here he explains why, inside Mastra, an agent is really just a workflow that loops. You'll learn why workflows first mattered when the models weren't good enough, why they're back as tasks get longer and more complex, and the three distinct aspects of durability — human-in-the-loop, resuming a multi-day agent run after a crash, and replaying a stream. Tony also demos a dynamic workflow that fans out ten agents in parallel to summarize every file in a codebase.
August 25, 2026
#109
The 0x-alpha Mystery, Qwen's 27B Beats Opus 4.8 & Legal Is AI's Breakout | This Week In AI
A shorter but juicy one. Shane and Abhi open on the debate everyone's having — do you want to use your agent to reach into someone else's tool, or do you want that tool to build the best agent and meet you where you already are? Then the week's biggest mystery: a stealth model called 0x-alpha that appeared on OpenRouter beating Fable and Sol on coding, with a wild one-hundred-trillion-tokens-a-day capacity that had everyone convinced it was a secret Google model — until someone made the server name itself, and it turned out to be Z.ai. From there: Qwen breaks the size curve with a 27B model that outranks Opus 4.8 on the agentic index and runs on your own hardware, open-weight tokens cross 62% on Vercel's AI gateway, and Salesforce ships Slack Code so humans and agents can work in the same channel. Plus the case that legal is AI's breakout use case — Harvey trained its own model, Tenet, on a Kimi K3 base, and legal is the fastest-growing Codex adopter, up 108x since February. In the quick hits: agents now burn 5x the tokens humans do, and Tobi Lütke vibe-codes a Git-at-scale binary over a weekend.
August 20, 2026
#108
The Three Ways AI Agents Fail in Production | Hendrix Liu, Respan
What does it actually take to run AI agents reliably in production? Hendrix Liu, co-founder of Respan (formerly Keywords AI, YC W24), joins Shane and Abhi to break it down. Respan is a one-stop LLM engineering platform — an AI gateway, observability, evals, and prompt optimization in one place. Hendrix shares the three failure modes every agent team runs into (silent failures, reliability outages, and runaway cost), and the techniques to catch each one: fallback routing across 1,000+ models when a provider goes down, online evals that flag hallucinations before a customer does, and spend limits at the per-key, per-model, and per-customer level. Plus the counterintuitive origin story, a live demo of their Scratch-style evaluators, and an honest take on build vs. buy for the LLM stack.
August 19, 2026
#107
Stripe Buys OpenRouter for $7B, New Models Drop & Anthropic Watermarks AI | This Week In AI
Money is moving, models are dropping, and GitHub's status page looks like a Christmas tree. Shane and Abhi open on the week's most-shared joke — the SF founder who realized his best friend was "emotionally attached to a cron job" — then get into Anthropic watermarking every bit of Claude's output to comply with the EU AI Act, and the reasoning-leak vulnerability that let researchers pull hidden thinking tokens out of frontier models. Then a wave of new models: Grok 4.6 lands right in the conversation with GPT-5.6 and Fable (and Elon is already teasing 4.7), GLM 5.3 pushes near-frontier for cyber defense, and Qwen ships an open-weights 27B you can actually run on your own hardware. On the money side: Stripe reportedly acquiring OpenRouter for $7B, Dynatrace buying Arize, CodeRabbit's $143M Series C (with an all-time-great Silicon Valley parody video), and Lovable raising $400M at $13.3B. Plus DeepSeek's new harness, Sol running ultra-fast on Cerebras, and the code-hosting land grab as Cursor Origin and Netlify both become Git hosts while GitHub keeps going down.
August 14, 2026
#106
Git Wasn't Built for Agents, So Mesa Rebuilt the File System | Agents Hour
Git wasn't built for AI agents — so Ben Warren's team rebuilt the file system from scratch. Ben Warren, co-founder and CTO of Mesa (YC S24), joins Shane and Abhi to explain why the file system is the missing primitive for AI agents, and how Mesa — a versioned, POSIX-compatible file system with Git-style branching — lets you run massively parallel agents without the pain of Git.
August 13, 2026
#105
What Are Agent Plugins, Are We Ready for AI Employees & OpenClaw Hacks a Gym | This Week In AI
An AI agent hacked a gym — and that's not even the biggest story of the week. Shane and Abhi open on Australia's first known autonomous AI cyberattack, where a Melbourne man's OpenClaw agent found a hole in a gym's booking API and canceled a stranger's reservation to move him up the waitlist. Nobody asked it to. From there: OpenAI's new Agent Plugins standard (built on Anthropic's own skills and MCP), the "AI employee" era with Lindy and Grok Bot, Meta opening the weights on Muse Glimmer, and OpenAI shipping GPT-5.6-Cyber while quietly postponing Astra after it hit "Critical" on cyber. Plus a run of acquisitions (Smithery to Arcade, ElectricSQL to Neon at Databricks, and Manus getting un-acquired by Meta), NVIDIA turning compute into a $500B asset class, and Anthropic signing a $9.1B, 20-year compute deal with a Bitcoin miner. In the quick hits: Claude proved more than 67% of the Riemann zeros lie on the line, the inside story of how OpenAI's agents used a JFrog artifact registry to coordinate a hacking spree, Unsloth Desktop, Sonnet 5's permanent price cut, Spotify's Xirp, Stagehand v4, and Harvey open-sourcing a synthetic law firm.
August 10, 2026
#104
Building the AI-Native Superhuman for Sales on Mastra - Cedar Copilot | Agents Hour
Isabelle Ilyia, co-founder and CTO of Cedar (YC W25), joins Shane and Abhi to demo Cedar Copilot — the AI-native "Superhuman for sales," built on Mastra from day one. Cedar started as an open-source framework for agentic frontends. Isabelle walks through the pivot: the same technology, verticalized for the messiest part of sales — the stretch from first demo to closed deal.
August 7, 2026
#103
OpenAI vs Anthropic: Who's More Dangerous? Is Google Giving Up? Open Weights Cook | This Week In AI
A delayed but stacked week. Shane and Abhi open on the question the whole industry is dancing around — is source code becoming assembly? — then dig into the strangest new status game in AI: OpenAI and Anthropic both admitting their models broke containment and hacked things, and treating it like a flex. Welcome to the felony bench, where the leaderboard is "whose model is most dangerous." Meanwhile the open-weight flood keeps rising — DeepSeek-V4-Flash, Qwen3.8-Max, and MiniMax-H3 all landing near frontier for a fraction of the price, to the point where defaulting to open models just makes sense. OpenAI slashes prices (Luna down 80%), making an already-rough week even rougher for Google: Jeff Dean leaves to found Discovery Loop, Demis Hassabis steps down to chair, and the read gets harder to avoid — is Google giving up on winning in AI? Plus OpenAI's "Apple is getting this wrong" clapback, Suno losing its copyright case, Meta's Muse Code, Ilya's SSI finally shipping, Prime Agent, OpenAI's Astra solving ten open math problems, and stacked PRs finally landing on GitHub.
August 4, 2026
#102
Freestyle's CEO Runs His Cloud From a House — Ben Swerdlow on Agent VMs
Ben Swerdlow is CEO and co-founder of Freestyle, which builds Linux VMs for AI agents. He joins Abhi Aiyer to explain why he refuses to call them sandboxes. Freestyle runs on bare metal in a house. There is duplicated internet, a battery, a backup kerosene generator, a fingerprint-protected door, and a hole cut in a wall. The move followed a $45,000 cloud bill and the discovery that an old computer outperformed the cloud version of an early prototype. Ben's advice to most of the audience is not to use his product. But he draws the line at execution time and names the case where a sandbox is simply the better tool. You'll also learn why Freestyle moved off Firecracker and what rust-vmm makes possible, the destructive-snapshot problem that caused Ben real pain, and why stripped-down kernels break on Supabase.
July 30, 2026
#101
Opus 5 vs Fable, Rogue Model Hacks Hugging Face & the Open Weights Letter | This Week In AI
Big week. OpenAI had a security incident, where a model reportedly broke out of its evaluation parameters and hacked Hugging Face — and the guardrails on OpenAI's and Anthropic's own models blocked responders from investigating, so they had to fall back to an open model to get to the bottom of it. Opus 5 lands close to Fable at 26% lower cost and tops the front-end arena, but the reviews are mixed, and the token usage is brutal — one chart clocks it at 97 million tokens on a task GPT 5.6 did in 4.3 million. Then the fight everyone's watching: NVIDIA, OpenAI, and a hundred others sign an open-models letter, Anthropic declines and calls for mandatory safety testing, and the room asks the obvious question — who gets to write the test? Plus MCP goes fully stateless (it's basically a REST API now), Kimi K3 weights drop with a K3.1 leak, SSI partners with NVIDIA, Stripe eyes OpenRouter for $10B, Cognition buys Poke, and DoorDash announces drone delivery — so your agent really can order you pizza by helicopter.
