Agents Hour

AI agents write code faster than CI can check it. GitHub reports a 4x jump in Actions minutes since 2023, and Anthropic says CI became its biggest bottleneck after job volume grew 25x in six months.

Daniel Worku, co-founder and CTO of StarSling, joins Shane and Abhi to talk about what CI looks like when agents write most of the code. Faster hardware helps, but StarSling also runs a background agent, built on Mastra, that A/B tests changes to your CI and opens a PR when it finds a speedup.

Daniel explains why agents make CI so expensive: more PRs, piles of low-value unit tests, and stacked PRs that rerun everything. He also makes the case that most unit tests aren't worth running, and that CI as we know it may not survive.

You'll also hear how Mastra's own CI went from half-hour waits to a fraction of that, and why code review is the next bottleneck. StarSling's new Review Runners let you bring your own model and run reviews on the same infrastructure as your CI.

Recorded live on AI Agents Hour, a weekly livestream by Mastra CPO Shane Thomas and CTO Abhi Aiyer. Mondays 12PM Pacific.

👤 ABOUT DANIEL

Daniel on X: https://x.com/dbworku

📚 MASTRA RESOURCES

WHAT IS MASTRA?

⏱️ CHAPTERS

Anthropic's IPO prospectus leaked, showing a $2 trillion valuation target on $4.6 billion in revenue, with a quarter of that coming from two customers. Google is back in the frontier race with Gemini 4 Argon, which tops the Vals Index and can output up to a million tokens at once. Decision models are suddenly everywhere: OpenAI's Decisions API, OpenRouter's new category, Cloudflare's Clef, GLiDE, Respan's Span-01 and Kev.

The White House got the big labs to sign an accord on superintelligence, Factory fired an advisor who then became Cognition's CRO, and Vinod Khosla picked a side in public. Shane and Abhi also talk about why maintainers like Sindre Sorhus and Hono are closing external PRs, and go through quick hits including Turso joining Supabase, OpenDots, Conductor Mobile, Karpathy on explainer videos, Claude Code mods, and a video model that passes the Turing test.

AI Agents Hour is a weekly livestream by Mastra CPO Shane Thomas and CTO Abhi Aiyer. Mondays 12PM Pacific.

📚 READ MORE

📚 MASTRA RESOURCES

WHAT IS MASTRA?

⏱️ CHAPTERS

OpenAI DevDay 2026 happened, so Shane and Abhi jumped on a special live episode of Agents Hour to go through everything that was announced and share their takes.

The big launch is Dots, OpenAI's always-on personal agent, which arrives right after Grok Bot and Meta's Muse. There's also GPT-6.1 Sol, a speed tier called Ultrafast, Codex Security Cloud, a Decisions API that goes after Jev, computer use in the Agents API, and a lot of new ChatGPT features (Spaces, Pages, slides, meetings) that compete with Google Docs, Notion and a long list of startups. The $200 Pro plan came back with half the usage, next to a new $500 tier.

They also talk about choosing your lock-in, what "Sign in with ChatGPT" means for Devin, and the funniest reactions online: the Dots demo flop and xAI buying dot.com to redirect it to Grok Bot.

AI Agents Hour is a weekly livestream by Mastra CPO Shane Thomas and CTO Abhi Aiyer. Mondays 12PM Pacific.

📚 READ MORE

📚 MASTRA RESOURCES

WHAT IS MASTRA?

⏱️ CHAPTERS

Anthropic might be back. Opus 5.5 performs at Fable 5.1 level for 40% less, and Sonnet 5.5 dropped the same day we went live, beating Opus on Terminal-Bench 4. OpenAI answered with GPT-6 Sol and Luna and a rumored long-running agent called Aeon. Meta launched Muse at Connect, then hired MongoDB's CEO to lead a new enterprise platform.

Welcome back to Builders Learn ML with Professor Andy! In this episode we're talking about training your own AI models.

Should you train your own model? Maybe. But probably not yet. If you haven't nailed your prompts, workflows and evals, training won't fix your agent, and it only starts to pay off once you're running millions of tokens a day. In this episode, you'll learn when it's worth it and how to do it.

There are two main techniques for training your own AI models: supervised fine-tuning (SFT) and reinforcement learning (RL). SFT teaches a smaller, cheaper model to copy a stronger one using thousands of good examples from your agent. RL gives an open-weight model like Qwen, GLM or Kimi a task and rewards it when it gets a good result, so it figures out its own way there.

The hard part of RL is the reward. Score it badly, and you get reward hacking, where the model learns that doing nothing is the safest way to avoid a penalty.

They also talk about RL environments and Harbor, unit tests vs LLM judges as verifiers, how much data you need, why good evals are worth having even if you never train, and where to start learning on your own.

Andy Lyu is the co-founder and CTO of Osmosis (YC W25), a platform that helps companies train their own models with reinforcement learning. Before Osmosis he was a tech lead on TikTok's real-time recommendations and data infrastructure team.

Recorded live on AI Agents Hour, a weekly livestream by Mastra CPO Shane Thomas and CTO Abhi Aiyer. Mondays 12PM Pacific.

📚 ABOUT OSMOSIS

📚 MASTRA RESOURCES

WHAT IS MASTRA?

⏱️ CHAPTERS

This week, X could not stop talking about Jev, a viral new "decision model" that hit 38 million views on launch. Shane and Abhi break down what it actually is, and Abhi builds with it live in seven demos. Jev is not an LLM. Instead of returning text, it returns structured output with a calibrated probability, trained with a new method its creators call RLCD (reinforcement learning for calibrated decisions). The pitch is 20 to 200x faster and 40 to 400x cheaper than an LLM, with output tokens free, and the community went wild: people cloned it, built programming languages and generative UI on top of it, and half of X argued over whether it is really just an if-statement.

Abhi spent the weekend adding Jev to every primitive in Mastra as a new classifier and assessed where it helps and where it does not. He wires classifiers into workflow branching, turns PII and abuse detection into cheap guardrail processors, tries tool-search pre-selection, tests model routing (he is not convinced), and shows tool approvals and classifier-as-judge scoring. The honest takeaway: it is a genuinely useful new tool for the belt, not the revolution the hype claims, and the marketing was better than the breakthrough.

Before all that, Shane opens on Gergely Orosz's deep dive into OpenAI's software factory and how Codex took over there, and why the reaction from a lot of people was "that's it?" Plus the first SF Software Factory meetup on October 14th.

Quick hits: Claude Code finally reads AGENTS.md, an open-weight 7B (Qwen-Image-2.1) beats Nano Banana 2.0, shadcn/lint gives agents a design-system linter, an offline OS hallucinated with Qwen 3.8, Factory raised $200M at a $5B valuation, and Tobi coins "slop grenades."

AI Agents Hour is a weekly livestream by Mastra CPO Shane Thomas and CTO Abhi Aiyer. Mondays 12PM Pacific.

📚 READ MORE

📚 MASTRA RESOURCES

WHAT IS MASTRA?

⏱️ CHAPTERS

This week AI doom went mainstream, and Shane and Abhi spend most of the show untangling it. Live from San Diego, they trace how one post racked up 171 million views plus a WSJ exclusive that ran before the post itself, kicking off the biggest safety fight in a while.

The spark was Jacob Coxon's resignation, amplified when Anthropic's own alignment science lead, Evan Hubinger, believes there's a greater than 10% chance AI kills everyone within a decade. Parker Thayer called the whole thing a well-funded PR operation. Then Dario Amodei published "We Must Pace the Frontier," arguing the industry should slow down, with a three-part plan Anthropic is committing to the first step of. Abhi traces how Dario's views got here, back to his OpenAI days and the ASL safety framework, so the "doomer" label lands with more context than the takes on X.

From there it's a pile-on: Elon says Dario was right, Hugging Face launches an Open Alignment Initiative, Sam Altman says OpenAI will do the same, David Sacks tells the "duopoly" to go ahead and slow down, and Zuckerberg says move faster. Meanwhile China's Z.ai raises $5 billion for a self-training system, which raises the obvious question: can you pace the frontier if everyone else is sprinting?

Plus Anthropic's threat report on how Claude has been misused (a Russian state group, ShinyHunters, a Mali system tracking 25 million SIM cards), and quick hits: GPT-Live-1, Cursor Projects, DeepSeek-V4.1-Flash, Devin Voice, Tailwind joining Shopify, and MCP agent skills.

AI Agents Hour is a weekly livestream by Mastra CPO Shane Thomas and CTO Abhi Aiyer. Mondays 12PM Pacific.

📚 READ MORE

📚 MASTRA RESOURCES

WHAT IS MASTRA?

⏱️ CHAPTERS

Tyler Barnes is a founding engineer at Mastra and the person behind its memory systems. In this chat with Shane, he explains observational memory - Mastra's fix for the thing every coding agent does where it fills up its context, compacts, and forgets what you were doing.

GPT-6 Astra is finally here, and it changes what a model launch even looks like.

Shane and Abhi open on the trailer for "Artificial" (Andrew Garfield as Sam Altman, in theaters Christmas), then get into the launch everyone's talking about: Astra's computer-use leap. This isn't just a coding model. It'll drive Blender, CAD, Notion, even Microsoft Paint, pulling a whole new class of technical-but-not-coding users into the fold. Abhi's been daily-driving it as a genuinely great collaborator; Shane's more mixed on the raw coding.

Then the benchmark wars: ARC-AGI-3 leaps from 7.8% to 98.6% in one drop, the AI-2027 curve suddenly holds, and Astra even tops Vending-Bench while being "more ethical" than Claude. Anthropic rushed Fable 5.1 out the door on September 1st, and Astra promptly sucked the oxygen out of the room. All this progress reignites the "should we pause AI?" debate, with Bernie Sanders in all caps, Austen and Dwarkesh going back and forth, and the EU deciding ChatGPT is a search engine. Plus OpenAI's agents crack the Navier-Stokes Millennium Prize problem (with a "significantly more capable than Astra" tease), Cognition raises $2B at $48B, Gemini 3.8 Flash, Qwen3.8-Max tops Code Arena, Claude "open sources" Commerce Agents, Runway Solaris, and TimesFM-3.

AI Agents Hour is a weekly livestream by Mastra CPO Shane Thomas and CTO Abhi Aiyer. Mondays 12PM Pacific.

📚 READ MORE

📚 MASTRA RESOURCES

WHAT IS MASTRA?

⏱️ CHAPTERS

"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.

Recorded live on AI Agents Hour, a weekly livestream by Mastra CPO Shane Thomas and CTO Abhi Aiyer. Mondays at 12 PM Pacific.

📚 ABOUT CONDUCTOR

📚 MASTRA RESOURCES

WHAT IS MASTRA?

⏱️ CHAPTERS