1022 | Agents Everywhere: This Week's AI Launches

Mia: Welcome back to the show, everyone. I'm Mia.

Milo: And I'm Milo. Today we've got a whole stack of launches that, honestly, all tell one story: AI agents are growing up. There's a model built just for agent decisions that claims to be dramatically cheaper and faster, marketing and support agents that actually do the work, and then the infrastructure layer — agents talking to each other, spending money, sitting behind approval gates.

Mia: Yeah, and it ends on the fun side too — Mac utilities, a Blender add-on, some design tools. Let's start with the thing that could change the economics of everything else we talk about today. Mouse-1. It's a decision model built specifically for agents, and the claim is eye-popping: 163 times cheaper and 65 times faster than GPT-6 Sol, returning probabilities in about 100 milliseconds.

Milo: Those are maker claims, so let's be careful — nobody's independently verified that on real workloads. But think about what it would mean if it holds up. Agents make thousands of tiny decisions. Should this email go here or there? Should this request be approved or escalated? Which route should this task take? Right now, a lot of teams pay big-model prices for those micro-decisions, or they fall back to brittle rules.

Mia: Right, and that's where the use cases matter. Routing, approvals, triage — high-frequency decisions where you need an answer fast and cheap, and a probability rather than an essay. If a model returns calibrated probabilities in 100 milliseconds, you could put a decision layer in front of every action an agent takes.

Milo: The open question, though, is accuracy outside of whatever benchmark they picked. Marketing tasks, say, versus messy real-world workloads. Is it actually as good on the decisions that matter, or is it only cheap because it's only good at narrow things? That's the gap between an impressive number and a product.

Mia: And it connects directly to what Notion is doing, because Notion clearly thinks cost is the bottleneck too. Notion Auto — agents that can run for up to 24 hours, and a new Auto router that gets results close to Opus 5.5 while costing roughly 90 percent less.

Milo: So the pattern is the same from two directions. Mouse-1 is a purpose-built decision model for the agent layer. Notion is routing work to cheaper models when the expensive one isn't needed, keeping quality near the top tier. Both are attacking the same problem: agents that run for hours at big-model prices don't pencil out.

Mia: Notion also introduced a new credits system, which matters if you're running long agent sessions — you want predictable costs when something chugs along for 24 hours. So whether it's smarter routing or specialized small models, the whole industry is converging on "make the agent loop affordable."

Milo: Okay, so if agents become cheap enough to run constantly, what are they actually going to do all day? That's where the next batch comes in — and marketing is the obvious first target.

Mia: SuperCMO is the flagship example here. It's an open-source set of skills, Apache-2.0 licensed, that turns Claude, Cursor, or Codex into an AI chief marketing officer. It creates UGC-style ads, videos, and photos.

Milo: Open-source is the interesting part. That means anyone can inspect it, fork it, adapt it to their brand. It's not a black-box SaaS you rent — it's a capability you own. The claim is essentially: your coding agent already knows how to do things; here are the skills that make it do marketing.

Mia: The unknown, as always with these, is output quality. Maker demos always look great. The real test is whether a SuperCMO-generated ad campaign holds up next to what a human team with a real budget produces. And who is it for? Probably solo founders and small teams who can't afford an agency — for them, "decent ads for free" changes the math entirely.

Milo: And marketing agents don't just make content, they need to go find customers. That's Needle — an outbound engine that's native for agents, accessible via MCP. It does lead generation, enrichment, and messaging across LinkedIn, Instagram, and email.

Mia: So the workflow an agent couldn't touch before — find the leads, enrich them, actually send the outreach — is now an MCP tool call. The agent doesn't just write the email, it sends it. That's a real shift in what agents can do end to end.

Milo: Which raises the scale question: if an agent can message hundreds of people on LinkedIn and Instagram, who's checking the messages before they go out? There's a real risk of agents spamming at machine speed. Nothing in the description suggests a review step, so that's a meaningful limitation.

Mia: Support is the other function moving to agents, and Scout from Fini takes a more careful approach. It's an AI in Slack for support teams. It reads tickets and reviews, identifies patterns, and drafts fixes — but every single change requires human approval.

Milo: That's a deliberately different design philosophy. Needle assumes the agent acts; Scout assumes the agent recommends and a human confirms. For support, where a bad fix can affect thousands of customers, that approval gate probably isn't a limitation — it's the feature.

Mia: There's also AskDoozy in this space — an AI in Slack that actually executes tasks in real systems like Stripe, Postgres, and GitHub, with a sandbox, an action log, and one price per workspace. So Scout analyzes and drafts, AskDoozy executes. Two different answers to "how much should we let the agent do?"

Milo: Now, once you have marketing agents sending outreach and support agents touching tickets, they all need shared plumbing. That's the next theme: infrastructure for agents that work together, safely.

Mia: BAND is the most literal version of that. It lets coding agents — Claude Code, Codex, Copilot, OpenCode — collaborate and coordinate tasks with each other in shared rooms.

Milo: So instead of one agent working alone, you might have Codex and Claude Code working on the same project, dividing tasks, seeing what the other is doing. The open question is obvious: does cross-agent coordination actually work in practice, or do agents just get confused by each other's half-finished work? Agents are bad enough at context management solo.

Mia: The other direction is governance. PrivOS is a self-hosted corporate AI operating system, free up to 10 users, where agents run in isolated containers connected via MCP.

Milo: Self-hosted and isolated is the enterprise answer to "can we trust these things?" If an agent misbehaves, it's contained. If it needs to reach a system, it goes through a controlled MCP connection. It's the difference between giving an intern the keys to the building and giving them a supervised desk.

Mia: And then there's knowledge. The Brains gives you 37 MCP servers of founder and investor knowledge — Paul Graham, Naval, that crowd — and every answer comes with a citation and a link. It's free and MIT licensed.

Milo: The always-cited part is what makes it more than a gimmick. If you're asking "what would Paul Graham say about this pricing decision," getting a quote with a link means you can check it. Without citations it's just an AI doing an impression.

Mia: It also matters for founders building agents — decisions like whether to gate features or charge from day one are exactly where a cited knowledge base could help. And speaking of money and access: agents need channels and budgets, which brings us to the control-panel layer.

Milo: Intelli is a plugin in the ChatGPT directory and an MCP for Claude, with 89 tools covering WhatsApp, Instagram, and website chat — and it confirms before it acts.

Mia: So think of it as the agent's hands for customer channels. It can reply on WhatsApp, handle Instagram messages, work the site chat — but with pre-action confirmation, so a human sees what's about to happen. 89 tools is a lot of surface area, though — more tools means more ways for something to go sideways, which is exactly why that confirmation step matters.

Milo: And if agents can send messages, they can also spend money — which is where Flyweil comes in. Virtual cards with spend limits and credit for ads, plus cost auditing via MCP so AI agents can be supervised financially.

Mia: This might be my favorite idea of the episode. You give your marketing agent a Flyweil card with a $50-a-day limit, it buys ad placements, and you can audit every cost through MCP. It's the same guardrail logic as a corporate card for employees, applied to agents. The trust question is still open, though — when an agent touches payments, who's liable when it overspends or buys the wrong thing? The spending limit answers part of that, not all of it.

Milo: Tend approaches trust from the personal side. It's a personal AI over iMessage and SMS that acts on your actual accounts — Gmail, Shopify, QuickBooks — but nothing acts without your approval, and it keeps an editable memory called the Brain.

Mia: The editable memory is underrated. You can go in and correct what it thinks about you and your business. That's very different from a black-box assistant accumulating assumptions about your finances.

Milo: And all these agents need to know what happened in your meetings. Twin for Desktop transcribes meetings and dictation on Mac and Windows, in over 140 languages, with a floating microphone and an MCP so agents can access what was said.

Mia: So the transcript becomes agent-readable context. Your agent knows what you promised the client because Twin captured the call. Notice the shape of all of this — channels, money, memory, meetings — the desktop is becoming the agent's cockpit, with a human sitting at the controls.

Milo: Which brings us to the Mac developer tools, and this segment is genuinely fun. Burrow gives you free, open-source virtual machines — MIT licensed — on Apple-silicon Macs, running Windows 11, Linux, and macOS with one click.

Mia: One click is the whole pitch. Virtualization on Apple silicon has been awkward; Burrow wraps it so a developer who needs to test on Windows or Linux just... does it. Free and open-source on top of that.

Milo: ilha solves a smaller but very real pain: it shows your Claude Code, Codex, and Grok Build usage limits right in the notch of your Mac, and automatically switches accounts before you hit a cap.

Mia: That's a quality-of-life tool born from a real frustration — getting rate-limited mid-task. Knowing your budget and switching before the wall is the kind of thing that seems trivial until you've lost an afternoon to it.

Milo: And then Agent Pets, which is exactly what it sounds like — every Claude Code or OpenCode session becomes an animal in Chrome, and you can approve agent actions from any tab. Six dollars, one time.

Mia: It's cute, but it's solving something real: agent approvals require attention, and sticking them in a persistent little pet means you don't have to keep a terminal in focus. Fun packaging, serious function. And hibi fits the same desktop-writing world — an open-source notes app, an Obsidian alternative, plain files in local folders, supporting markdown, Typst, and LaTeX. It's in alpha, so early days.

Milo: Koi belongs here too — a Mac app that's three tools in one: offline dictation with Fn-Option, an AI that can see your screen with Fn-Cmd, and call recording with no button, Fn-Control. Lifetime pricing starting at 59 dollars. And iglass is an open-source Liquid Glass theme for iTerm2, Ghostty, and kitty, using real glass on macOS 26, with contrasted colors, and it's reversible via uninstall.

Mia: Small utilities, one-time prices, open source where possible. The one caveat across this whole segment: these are tiny teams making tiny tools, and long-term support is the honest unknown. A six-dollar pet or a free theme doesn't fund years of maintenance.

Milo: Okay, last stretch — from design to delivery. This is about shipping things that are actually production-ready. UseIt Studio is an AI design agent, and its differentiator is that it produces real, editable files with dimensions, bleed, and dielines — ready for production.

Mia: That's the gap it's attacking. Most AI image tools give you a pretty picture; a print shop needs exact dimensions, bleed margins, and dielines. If UseIt genuinely delivers files a printer accepts without back-and-forth, it changes who can make physical products.

Milo: Then NativeShip, on the integration side: it creates and publishes a native n8n integration from your SaaS's existing API, with a preview, its own GitHub and npm presence, and traceable versioning.

Mia: Every SaaS eventually gets asked "do you have an n8n integration?" and it's usually weeks of work. NativeShip automates that from the API you already have, and the versioning matters because integrations rot when the underlying API changes.

Milo: And once you ship, you need to know if it works — Flagsmith Experiments adds open-source A/B testing on top of feature flags you may already have, with a Bayesian engine, running on your own ClickHouse or Databricks warehouse or fully managed.

Mia: Building experiments on existing flags is smart — you've already got the toggle infrastructure, this adds the statistics layer. And the option to keep data in your own warehouse is the enterprise-friendly move. The open question for all three: adoption versus incumbents. Established design tools and experiment platforms have gravity.

Milo: Quickly, the adjacent ones rounding this out. Lea by LeanAIOS is an AI platform for construction with 18 specialist agents, takeoff, FIDIC deadlines, and 3D testing, with unlimited users. Animatica Marionette is a Blender add-on, 59 dollars, that lets you pose human rigs by dragging hands and feet, running offline on CPU in about 15 milliseconds per pose.

Mia: Hotly is open-source premium messaging — one creator with 536K followers reportedly made $22,400 in 60 days using it — note that's a maker-reported figure — and you can self-host with zero percent commission. Yarite is an iPhone app with no account and no server that remembers contacts through six circles, free up to 15 people, then a one-time $9.99.

Milo: Nimbalyst Wiki is a local plugin that generates a markdown wiki from your coding agents — Claude Code, Codex and others — recording decisions and open questions as it goes. Blooio is a provider for Apple Messages for Business with a unified API, Apple Pay, forms, plus RCS, SMS, and WhatsApp support. And Querri Library organizes data around the questions your business actually asks, flagging alternative interpretations in 73 percent of answers versus 12 percent for general AI.

Mia: That Querri number is interesting — surfacing when data could be read a different way is exactly what stops bad decisions. Alright, that's the full sweep.

Milo: So if you take one thread away: cheaper decisions, agents doing real work in marketing and support, infrastructure to coordinate and constrain them, control panels for money and channels, and tools that make the whole thing feel human-scale — on a Mac, in a notch, or as a pet in your browser.

Mia: The caveat to carry with you: most of the bold numbers today are maker claims. The 163x cheaper model, the $22K in 60 days — treat them as promises until you see them on your own workload.

Milo: Exactly. Thanks for listening, everyone — we'll see you next time.

Mia: Bye, everyone.