0902 | Agents Everywhere: From Canvas to Cloud

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

Milo: And I'm Milo. Today we're doing something a little different — a full briefing of yesterday's launches, all of them from one pool of announcements, and we've tried to pick apart which ones actually solve a clear problem for a real person, and which ones are just... shiny.

Mia: Right, and fair warning before we start: everything we know comes from the makers' own descriptions. We haven't tested any of this. So when you hear a benchmark number or a claim, treat it as a claim, not a verified result. That said, there are some genuinely interesting patterns today.

Milo: The big one being: AI agents are stops being chat windows and start being... coworkers with desks. Persistent ones. Let's start there.

Mia: Yeah, let's talk about the workspace story, because three launches basically triangulate the same idea from different angles. The first is called Nodeterm, and the pitch is: your terminal and your coding agents become nodes on an infinite canvas. And the part the maker emphasizes most is persistence — tmux sessions that survive a full reboot.

Milo: Which sounds mundane until you've lost a three-hour agent run because your laptop slept. The claim is you can restart your machine and come back to everything running. And it also supports resuming over SSH, or from your phone.

Mia: That phone part matters, actually. It means the agent keeps working while you're away, and you check in from anywhere. It's a very different mental model from "I opened ChatGPT, I asked something, I closed the tab."

Milo: But Nodeterm is still fundamentally your machine. The second launch, Murmell, pushes it further — it's a shared cloud canvas for agents. Unified agent memory across the canvas, and the claim is that when you close the lid and come back, the agents' state is restored precisely where it was.

Mia: And there's a genuinely thoughtful detail: file reservations. If two agents want the same file, one can book it, so they don't trample each other. That's a real coordination problem once you run several agents at once, and most tools just... don't address it. It also supports real-time collaboration, so it reads like Google Docs but for agent work.

Milo: The open question with both of those is reliability at scale. A demo where two agents share a canvas is one thing; running this for weeks with real workloads is another. And neither launch gave us pricing, which is always a signal to be a little careful.

Mia: The third piece of this triangle is Superset Mobile — running coding agents in parallel from your phone. You launch tasks, watch progress, review diffs, and it pings you when something needs attention, syncing with your desktop or remote workspaces.

Milo: So Nodeterm says "your terminal is persistent," Murmell says "agents share a brain and a canvas," Superset says "you supervise them from the bus." Put together, the direction is pretty clear: agents as long-lived, device-agnostic coworkers rather than sessions you babysit.

Mia: And there's one more piece that fits here — Cosmic Agent Plugins. The idea is that a plugin bundles a vendor's official MCP server, plus skills, plus a token slot, into a single agent. So instead of wiring up integrations yourself, the agent gets real tools — GitHub, Stripe, the maker says others are already live.

Milo: That's the "hands" to the workspace's "desk." If your agent is going to live on a canvas and keep memory, it needs vetted, official tools rather than whatever someone stitched together. Again: claims, not verified. But the shape of the story holds together.

Mia: Okay, staying in the agent world but moving closer to the actual developer — coding agents going native and parallel.

Milo: The headline here is Kilo Code for JetBrains. Open-source, native plugin, and the three things the maker leads with are: parallel agents running on isolated worktrees, GitHub PR and diff review inside the IDE, and over 500 models to choose from.

Mia: The worktree detail is the one I'd underline. If you run three agents in the same checkout, they'll destroy each other's changes. Isolated worktrees is how you actually make parallel agents safe, and it's the same instinct as Murmell's file reservations — this coordination problem is showing up everywhere.

Milo: And "500+ models" means you're not locked to one vendor's agent experience — you can swap models per task. Open-source also means you can read what it does. Though open-source doesn't automatically mean good; it means auditable.

Mia: Then there's Naseem — a macOS agent written natively in Swift, no Electron. It works with files, the terminal, and the iOS simulator, and it uses the maker's own model. That's a notable bet: native performance and a proprietary model, in a market where everyone else wraps existing APIs.

Milo: The no-Electron part will matter to Mac users who are tired of every tool being a Chrome instance in disguise. But the proprietary model is a black box — you can't verify capability claims about it the way you could with, say, Kilo Code pointing at known models.

Mia: And the wildcard in this group is Sider Code, from the Sider browser extension. You describe what you want in natural language, and the AI reads the page and writes code to change it. Saved features can be toggled on and off. Free for the first three days after launch.

Milo: That's a different audience, right? Kilo and Naseem are for developers in their tools; Sider is "any webpage becomes yours to remix." Interesting, and the toggle-to-save is a nice touch, but it also raises the obvious question of what happens when the generated code conflicts with the site's own updates. Not addressed in the launch.

Mia: So across all three: coding agents are leaving the chat sidebar and embedding into the actual environments — IDE, OS, browser. And going parallel.

Milo: Which sets up the next theme nicely, because if agents are going to act in the world, one of the biggest places they act is... talking to your customers. Voice and messaging agents.

Mia: This might be the most commercially mature group of the day, because these have real pricing attached. ThunderPhone is a self-serve platform for AI phone agents, at two cents a minute including the model cost.

Milo: Two cents a minute is a concrete, comparable number — that's the kind of claim you can actually evaluate against a human call center. The maker claims 99.4% on Big Bench Audio in the Storm tier, across 47 languages, using multi-model fusion to improve accuracy.

Mia: And we should be careful — Big Bench Audio is a benchmark, and benchmark performance doesn't tell you how it handles an angry customer at 2 a.m. But the fusion approach — multiple models checking each other — is at least a plausible mechanism for the claim, and the pricing transparency is a good sign.

Milo: Then Keiki, which is the messaging-side version: one customer-facing agent that runs across SMS, iMessage, WhatsApp, Slack, Telegram, and email, with memory, tools, and — importantly — human approval mechanisms.

Mia: Human approval keeps coming up in this group, and it's the trust pattern. Tovel AI is the third piece: it listens to your calls and turns them into CRM records, to-dos, and follow-up email drafts. Deeply integrated, so no Zapier glue. And nothing sends without approval. There's also companion hardware, preselling at 99 dollars.

Milo: So the pattern across all three: AI does the drafting, the transcribing, the calling — but a human signs off before anything reaches a customer. That's a sensible on-ramp, and honestly it's probably where most businesses need to start. The open question is whether the approval step stays valuable or becomes a rubber stamp people stop reading.

Mia: And one caveat on Tovel — we know the hardware presale price, but not how the software itself is priced long-term. That's worth checking before committing.

Milo: Alright, from agents talking to people, to AI making things and teaching things. This group is about full content generation.

Mia: Impractical is the boldest pitch: AI motion design where an agent, connected through MCP, generates what the maker describes as After Effects–quality finished video. Prompt to finished piece in about twenty minutes, and you revise by just saying what to change in natural language.

Milo: Twenty minutes to a publishable motion graphics piece — if true, that changes who can make this kind of content. But "AE-quality" is the maker's own claim, and anyone who's worked in After Effects knows the gap between "looks like a motion graphics video" and "actually matches a professional's standards." We'd want to see real output.

Mia: Then Creatium Coach, which comes with the most interesting number of the day. It's a multimedia AI coach — you pick something like an "Aristotle world," and it generates simulations and role-plays toward your goal. The maker cites research suggesting an effect size of g equals 0.48, versus an average of 0.37 for in-person instruction.

Milo: That's a real research citation, not just marketing copy — though a caveat: effect sizes from studies don't automatically transfer to a specific product. The claim is that the product builds on that evidence, not that the product itself has been tested at that level. Still, it's more substance than most launches offer.

Mia: And Gauth AI Course takes the teaching idea to its simplest form: any subject becomes an interactive course instantly. It ships with over 200 built-in US high school math lessons, from Algebra 1 through AP Calculus, with in-lesson quizzes and an AI tutor you can interrupt mid-explanation to ask questions.

Milo: The interruptible tutor is the detail I like — most video courses are one-directional. If you can stop and ask "wait, why did that step work," that's closer to actual tutoring. Again, quality of the explanations is the thing we can't verify from a launch page.

Mia: So the theme: AI moving from assisting humans who make content, to generating and teaching the content itself. And music belongs in that conversation too — Happy Shrimp, from Alibaba. You give it an idea, it generates a complete song: lyrics, melody, arrangement, vocals. You can bring your own lyrics or ask for instrumental only.

Milo: The fact that a major company is shipping this tells you generated music is going mainstream. What we don't know from the launch is licensing, quality benchmarks, or pricing — so, open questions as usual.

Mia: Okay — from software that teaches, to actual physical hardware. And this pair could not be more different.

Milo: On one end, HONOR's Robot Phone. A flagship phone with a 4-degree-of-freedom titanium gimbal built into the body, physically moving a 200-megapixel main sensor. The gimbal handles stabilization, subject tracking, and autonomous shooting — the camera moves on its own. And the whole thing is 9.6 millimeters thick.

Mia: A moving camera inside a phone at under 10mm is a serious engineering claim, and if it works, it changes what phone video looks like — no more hoping your hands are steady, and the phone can frame and follow a subject while you do something else. But "autonomous shooting" raises questions: how well does it track, how does the moving sensor handle dust and drops, and what does it cost? None of that is in the launch.

Milo: And on the completely opposite end of the seriousness spectrum: Notchling. A little creature that lives in your MacBook's notch and holds files you drag onto it. Free, under one megabyte, and nothing leaves your Mac.

Mia: Which is honestly a charming counterpoint. One company builds a gimbal robot phone; one developer builds a digital pet that holds your downloads. Both are answers to "how should hardware and software relate physically" — just from different universes.

Milo: And Notchling's privacy claim — nothing leaves the Mac — is the kind of thing you can actually check with a tiny open app. Low stakes, easy to verify, kind of delightful.

Mia: Right, and that privacy theme actually carries into our next group: infrastructure, transparency, and trust as products.

Milo: This might be the most quietly important section of the day. Start with BobVault — a CLI backup tool built together with AWS. The pitch: your source code, configs, and AI context get encrypted locally, then uploaded. Keys never leave your machine. The cloud only ever holds ciphertext.

Mia: "Built with AWS" is notable — a cloud vendor acknowledging that some customers want zero-knowledge backup is a shift. The architecture claim is sound in principle: if the key never leaves your machine, the provider can't read your data. The usual caveat with zero-knowledge is on you, though: lose the key, lose the backup. That's the tradeoff.

Milo: Then TrustedRouter, which attacks trust from a different angle. It's an OpenAI-compatible API giving you 600-plus models from 90-plus providers, and the gateway is attested, keeps no prompt logs, and is open-source.

Mia: "Attested" is the key word — it means there's a verifiable claim about how the gateway runs, not just a promise in a blog post. If you're routing customer data through a model aggregator, log-free and auditable is exactly what you want to ask for. And open-source means the community can check the claims.

Milo: The third one is price transparency: CGI, the first open-source GPU price index. Data from 28 providers, refreshed every 15 minutes, with a verifiable methodology, in dollars per hour.

Mia: GPU pricing has been famously opaque — every provider quotes differently. A public index with a documented method doesn't just help you shop; it pressures the whole market toward honesty. The thing to watch is whether providers' listed prices match what customers actually pay.

Milo: And the last in this group is more developer-facing: EAS Observe, performance monitoring for Expo and React Native. It measures real startup time and per-screen usable time on actual devices, and marks each build and OTA update separately, so you can see if an update made things slower. Free tier is 100,000 events a month.

Mia: Real-device measurement matters because emulators lie. And per-OTA-update marking is smart — with over-the-air updates, your app's performance can silently change without any app store release, and now you'd see exactly which update caused it.

Milo: One more that fits the transparency discussion loosely: Sourclip 2.0, a Chrome extension for NotebookLM. Bulk capture, merging notebooks, export. Pro is 24 dollars a year. It's a productivity layer on someone else's product — which always carries the risk that the platform changes and breaks it. Cheap to try, though.

Mia: Alright, let's close with the grab bag — open source projects and consumer oddities, the stuff built for delight and focus.

Milo: ChannelOS first, because it's a genuinely fun idea: a local-first Windows app that turns your own media library into programmable cable TV channels. Channel guide, video on demand, gamepad support. MPL-2.0 licensed, open source.

Mia: The genius of it is the frame — you already own the movies and shows; this gives them the ritual of cable. Flip on, see a guide, something's already playing. Local-first means your library never leaves your machine, and open-source means you can verify that.

Milo: Then Folio, which is for people drowning in reading: it aggregates your open tabs, RSS feeds, and newsletters, and lays them out into a magazine with a table of contents, delivered automatically as PDF or EPUB to a reMarkable or Kindle. Five dollars a month.

Mia: That's solving a real workflow — the "I'll read this later" pile that never gets read. Moving it to an e-ink device changes the context from "sitting at my computer" to "reading." Five a month is reasonable if you actually use it; the test is whether you really do.

Milo: The nostalgia piece: nOS4, a full browser recreation of the iOS 4 experience. You can open apps and play games. The maker is upfront that it's a personal nostalgia experiment, not a product — which is honest, and honestly charming.

Mia: Two language and music tools round it out. WaseiGo teaches over a thousand wasei-eigo — Japanese words that look like English but mean something different. Fifty words free, then a one-time purchase. That's a real trap for English speakers in Japan, so the problem is genuine.

Milo: And ARC-24, an eight-track groovebox for iOS. Scale lock so you can't play a wrong note, plus synth, sampler, drum machine, and looping. Seven-day trial, then a one-time 9.99 dollars. No subscription.

Mia: One-time purchase in a music app in this economy is almost rebellious. Between that and WaseiGo's one-time purchase, there's a small but noticeable counter-trend of buy-it-once tools in this batch.

Milo: So let's pull the threads together. If yesterday tells us anything, it's that agents are becoming infrastructure — persistent, parallel, tool-using, spread across your terminal, your IDE, your phone, your customer's phone line.

Mia: And the interesting fault lines to watch: persistence and coordination — can these agent workspaces stay reliable at scale? The human-approval pattern in customer-facing agents — does it hold as a trust mechanism? And trust as a feature in general — zero-knowledge backups, attested gateways, open price indexes. Makers are learning that "trust me" is no longer a pitch.

Milo: And the honest caveats stand: nearly every capability claim and benchmark today comes from the makers themselves. We haven't run these tools, and pricing or availability details are missing in more than a few cases. Check before you build your week around any of them.

Mia: That's the briefing. Thanks for listening, everyone — we'll be back with the next round.

Milo: See you then.