1023 | Agents Everywhere: This Week in AI Tools

Mia: Welcome back to the show — I'm Mia, and today we're doing something a little different from a straight news rundown. We went through the launches that hit in the last twenty-four hours and picked a small set where the user problem is clear, the solution is actually differentiated, and the maker gave us enough concrete detail — capabilities, pricing, availability — to have a real conversation instead of just reading marketing copy.

Milo: And I'm Milo. One thing to set expectations before we dive in: everything we know about these products comes from the launch descriptions themselves. We haven't tested any of this, and where a maker claims results, we'll treat that as a claim, not a verified fact. But there is a genuinely interesting thread running through today's set: AI is moving from something you chat with into something that waits for you, runs locally, or sits inside the tools where your work already happens.

Mia: So let's start with the one that frames that thread best: Claude Duet from Anthropic. This is a proactive agent, meaning it works in the cloud on your behalf and only notifies you when it's finished something or when it hits a decision it needs you to make. It's in early access for Pro and Max users in the US.

Milo: That "only pings you at decision points" part is the whole design shift. Most agents today are things you babysit — you open a window, you prompt, you wait. Duet's pitch is more like delegating to a colleague: it goes away, does the work, and comes back either with a result or a question. The obvious unknown is how reliable that is at scale — how good is its judgment when it decides something on your own behalf? The announcement doesn't answer that, and we shouldn't pretend it does.

Mia: And notice what it contrasts with, because the next product is almost the mirror image. Bits is an open-source personal agent that runs locally on your machine. It keeps its memory in Markdown files, it's built to recover from crashes, and you can point it at whatever model you want.

Milo: So Claude Duet says "trust the cloud, check in when I'm done," and Bits says "keep the agent on my own hardware, keep the memory in files I can actually read and edit myself." The crash recovery piece is worth pausing on, too — that's a very practical admission that long-running agents fall over, and a local one needs to survive that. And being open source with any-model support means the trust model is completely different: you can inspect everything.

Mia: Right — one is a hosted, curated experience behind a subscription; the other is a hobbyist-friendly, self-hosted setup. They're answering the same question — where should an agent live — from opposite ends.

Milo: And there's a third answer in this same group: Concierge, from a company called Juniors AI. It's omnichannel customer support AI, but the twist is that it runs on the customer's own infrastructure — their Postgres, their Redis, their own LLM API keys — and they say there's no markup on top.

Mia: That last bit matters more than it sounds. A lot of support tools quietly resell you the underlying AI at a margin. Concierge's pitch is that you bring your own keys and they don't take a cut of your inference costs, plus your data never has to leave your stack. For companies with strict data residency or compliance needs, that's the entire reason to pick it.

Milo: So three products, three locations for the agent: Anthropic puts it in their cloud, Bits puts it on your laptop, Concierge puts it inside your own servers. And the shared idea across all three is that agents are becoming infrastructure — something that persists and keeps working — rather than a chat window you open and close.

Mia: Before we move on, it's fair to note what we don't know for any of them. Pricing beyond the headline access tiers isn't clear for Duet, we don't know how Bits holds up under real long-term use since it's community-driven open source, and Concierge's "no markup" claim is exactly that — a claim. We'll come back to this thread later in the show, because agents show up again in ad tools and shared inboxes.

Milo: Speaking of agents that have to work in real time — let's talk voice. The big launch here is ElevenLabs rolling out real-time speech translation: two-way, over ninety languages, and importantly, it keeps a consistent voice, plus it comes with subtitles and transcription.

Mia: The "consistent voice" detail is the differentiator. Real-time translation tools have existed in various forms, but usually the translated voice doesn't sound like the original speaker. If you can speak English and come out the other side sounding like yourself — same voice, another language — that changes live meetings, streams, customer calls. That's a much bigger deal than another subtitle tool.

Milo: And bundling transcription and subtitles makes it usable in contexts where you can't play audio out loud, or where you need a record of what was said. Of course, the open questions are the ones nobody's launch copy answers: what's the actual latency in a real call, how does it handle crosstalk or accents, how good is it in the less common of those ninety-plus languages? Those are field tests, not demo videos.

Mia: Now, in the same space but from a different direction: Mistral released Voxtral 3, a family of voice models. Transcription in over fifty languages, text-to-speech with emotion control, and a voice agent SDK built on LiveKit.

Milo: So ElevenLabs is shipping an end-user product, and Mistral is shipping the components — "here are the models and the SDK, build your own voice agent." The emotion-controlled TTS is interesting on its own; it's aimed at making synthetic speech less flat, which matters for anything customer-facing. And tying the SDK to LiveKit signals they're targeting real-time, conversational use cases specifically.

Mia: And then there's a third voice product that approaches the problem from an angle I really like: Kugur Ear 1. Instead of pretending transcription is always right, it flags only the words it's uncertain about, shows candidate alternatives, and lets an agent ask a confirming question. And it's priced at forty-eight euro cents per hour.

Milo: That pricing tells you they're going for volume — long recordings, always-on capture — and at that rate the cost of flagging uncertainty is basically negligible. But the design philosophy is the real story: most transcription tools silently commit to a best guess, and you find the errors later when it matters. Kugur makes uncertainty a first-class part of the output. An agent that asks "did you mean X or Y?" is honest in a way most transcription isn't.

Mia: So the voice story today has a shape: it's going real-time, it's going multilingual, and it's starting to be honest about what it doesn't know. What we still don't know, for any of these, is how they perform in the wild — latency and accuracy outside a controlled demo.

Milo: Which connects to the next group, because it's about media that's also generated in real time from simple inputs — video and 3D. Fastlane 2.0 is the most end-to-end of these: you give it your site URL, and it claims to handle brand learning, competitor research, generating AI video and image ads, delivering them on Meta, and then learning from the results.

Mia: That's the full advertising loop in one pipeline, at least as described. The interesting claim isn't any single step — AI ad creative exists, Meta delivery exists — it's the closed loop at the end, where the system learns from performance and adjusts. Whether that learning loop actually produces better ads over time is exactly the kind of thing we can't verify from a launch post.

Milo: Right, and note the trust asymmetry: you're handing over brand identity and ad spend to an automated pipeline. The maker says it learns your brand from the site; whether that produces on-brand creative or generic AI slop is an open question.

Mia: CutBay is a cousin of this — it's video generation through chat. You give it a prompt, a URL, a PDF, and it produces a finished video including the script, the narration, and the editing.

Milo: So Fastlane is "ads from your website," CutBay is "any video from any input, via a chat thread." What they share is that the input is no longer footage — it's intent. You describe or point at something, and the editing, scripting, and narration are all downstream of that.

Mia: And then Voxelo 3.0 takes it into 3D. You record about three minutes of video on your phone, and in roughly an hour it produces a photorealistic 3D twin — a digital replica of a person or object — and publishes its attributes as JSON-LD.

Milo: The JSON-LD part is the unglamorous but important bit. It means the 3D model isn't just a visual asset; it carries structured, machine-readable data about its properties. That's what would let a 3D twin plug into product catalogs, e-commerce, or search — the twin becomes data, not just a render.

Mia: So across Fastlane, CutBay, and Voxelo, the pattern is prompt-driven media: text or a phone clip in, finished asset out. And Dadan reinforces it from the creator side — it's one AI video workspace that combines recording, editing, meetings, and guides, and you can make videos interactive with polls and calls to action.

Milo: The common open question across all of them is quality control and authenticity. When anyone can generate a photoreal twin from a phone video or a finished ad campaign from a URL, the bottleneck shifts from production to verification. Who checks that the twin is accurate, or the ad is truthful? None of these launches addresses that, and it's the honest gap in the category.

Mia: And that "capture becomes machine input" idea is exactly where Loom is heading too, so let's stay on it. Loom has a new native Mac app, and the core move is that recording your screen and your voice becomes a structured prompt for agents — they name Claude, Cursor, and Rovo.

Milo: This is a real repositioning for screen recording. Loom's classic use case was asynchronous communication — "watch me explain this bug." Now the recording isn't just for a human to watch; it's structured input that an agent can act on. You talk through a problem on screen, and Cursor or Claude picks it up as a task.

Mia: And Dadan is doing something adjacent but different — it's less "recording as agent input" and more "recording as the center of a workflow." Its workspace combines the capture, the editing, the meeting, and the guide in one place, with interactivity layered on. So Loom feeds agents; Dadan keeps humans in the loop but consolidates the toolchain.

Milo: The shared bet is that captured media becomes machine-usable — either directly by agents, in Loom's case, or as the raw material of a unified workspace, in Dadan's. The unknown is the same for both: how well does anything downstream act on loosely captured context? A rambling three-minute screen recording contains a lot of noise; turning that into a reliable prompt or a polished guide is the hard part, and neither launch demonstrates it end to end.

Mia: Let's shift to inboxes and reminders, because that's where a lot of "AI inside your existing tools" is landing. Claire Mail is a Mac and iPhone app that merges Gmail, your calendar, and Apple Reminders into one place, with AI that handles inbox cleanup and gives you a morning summary.

Milo: The morning summary is the piece that could actually change a daily routine — instead of opening four apps and triaging yourself, you get one briefing that connects what's in your inbox with what's on your calendar. That's a concrete, checkable promise, at least in principle.

Mia: And Notif stays closer to a single problem: it layers onto Apple Reminders, adding context — what it calls "notice" information — and next-action buttons to each reminder. It's free, with a Pro tier at four dollars ninety-nine a month.

Milo: That's a nice small illustration of the same idea. A reminder is usually just a string of text; Notif turns it into something with context attached and an obvious thing to do next. Low price, single surface, but it's the same philosophy: the tool shouldn't just store your intent, it should help you act on it.

Mia: And then Resend Inboxes connects back to our very first topic. It creates shared inboxes with an API-first approach, so you can connect agents to them or build workflows on top.

Milo: Exactly — this is the infrastructure version of the inbox story. Claire Mail is AI triage for your personal inbox; Resend is saying "here's a programmable shared inbox where agents can live and work." If you remember Concierge from the top of the show, running support AI on your own infra — a shared inbox like this is the kind of surface that agent would operate through. So the themes are genuinely converging: agents need surfaces to work through, and inboxes are becoming those surfaces.

Mia: Which brings us to ads and lead generation, where that agent pattern shows up in a commercial context. Playad automates ad operations using multiple AI agents, each with a role — research, strategy, analysis, and bidding — starting at thirty-nine dollars a month.

Milo: The multi-agent structure is the differentiator here. Rather than one assistant doing everything, it splits the ad workflow the way a human team would: one agent researches, one plans, one analyzes, one manages bids. And the thirty-nine dollar entry price puts it in reach of small teams, not just agencies. Of course, "our agents run your ads" is a big claim, and the proof would be in sustained performance data we don't have.

Mia: Causo is the lead-gen counterpart. You describe your ideal customer in natural language, and it finds leads from publicly available information — with evidence attached to each lead. And notably, you can use it through ChatGPT or Claude via MCP.

Milo: The evidence part is what earns it a spot. Most lead tools hand you a list and you take their filtering on faith. Causo's pitch is that every lead comes with the reasoning — here's why this company matches your description, based on public data. That makes it auditable, which is exactly the right instinct in a category full of bad data.

Mia: And building it on MCP rather than as a standalone app is a smart distribution choice — it meets people inside the tools they already use. The open question is the perennial one in this category: data quality. Public data goes stale, and "evidence-backed" is only as good as the underlying sources.

Milo: And Fastlane belongs in this conversation too — it's in our supporting material here, because URL-in, ads-out is the same "specialized agent pipeline" idea applied to creative. So across Playad, Causo, and Fastlane, marketing is being decomposed into agent pipelines, each product claiming a slice of it. What none of them establishes, from a launch post alone, is ROI.

Mia: Now let's talk about tools that close loops — developer and file tools where the gap between "something went wrong" and "something got fixed" gets smaller. heal.dev runs your application in a sandbox that mimics production, and when it finds a bug or vulnerability, it reports it as reproducible, with a failing test attached.

Milo: The failing test is the whole point. A bug report that says "here's a description of a crash" still leaves an engineer hours of reproduction work. A failing test means the report is already executable — you fix the code until the test passes. And running in a production-like sandbox means the bugs it finds resemble real ones, not toy lab conditions.

Mia: Whiteboard does something similar for the design-to-code side. It's an open-source shared canvas where you can draw diagrams and see agent traces, and you can jump directly from those visuals into the code.

Milo: So the canvas isn't a dead-end drawing; it's a navigable layer over the actual code. And showing agent traces matters as agents write more of our software — being able to see what an agent did, visually, and click through to the code, is a form of accountability. Being open source helps that story too, since the community can inspect and extend it.

Mia: And DeepCell takes the loop-closing idea to documents. Its .deepcell format links a spreadsheet, a report, and slides together, so one edit updates all of them, and it exports to xlsx, docx, and pptx.

Milo: Anyone who's maintained "the same numbers" in a spreadsheet, a written report, and a slide deck knows exactly the pain point. The clever part is the export step — it doesn't force you out of the formats everyone already uses; it keeps them in sync underneath. Whether it handles every edge case of real-world documents is, as always, unproven, but the problem is real and clearly stated.

Mia: And Archivist fits here as a supporting example — you drop a folder of images and videos, it renames them according to a naming convention, it does collection management, and you can share with a single link.

Milo: It's the smallest of these tools but it's the same instinct: the gap between "I have a pile of files" and "my files are organized and shareable" is a loop that shouldn't require manual labor. So heal, Whiteboard, DeepCell, and Archivist are all attacking friction between states — between a bug and a fix, a diagram and code, a number and its copies, a folder and a shareable collection.

Mia: Two more, and they're about speed and filtering. Mercury 3 is a diffusion-based language model scaled forty times, claiming over a thousand tokens per second and outperforming GPT-6 Luna on benchmarks.

Milo: Diffusion LLMs generate text differently from the token-by-token models everyone knows — they refine output in parallel, which is where the speed comes from. A thousand tokens per second isn't just a benchmark number; if it holds up in real use, it changes what feels possible in interactive applications. But two big caveats: those benchmark comparisons are the maker's own claims, and pricing and availability aren't something we have details on.

Mia: And then Radaric, which sits on the consumption side. It's an AI feed reader that pulls together RSS, Telegram, X, and YouTube, and it learns your preferences from the reasons you approve or reject items.

Milo: The "learns from reasons" mechanism is the part worth dwelling on. Most recommendation systems infer taste from clicks, which rewards whatever's clickable. A filter that asks you why you rejected something and updates from that stated reason is a different model of preference — closer to how a human editor learns. Whether people actually take the time to give reasons is the practical question, but the design intent is genuinely different.

Mia: Which is a good place to land, actually. Put Mercury 3 next to Radaric and you get the full arc: generation is getting faster and cheaper, and curation is getting smarter to compensate. The future of consuming AI output isn't just better models — it's better filters between the flood and you.

Milo: That's the show. Everything we covered — Claude Duet, Bits, Concierge, ElevenLabs, Voxtral 3, Kugur Ear 1, Fastlane, CutBay, Voxelo, Loom, Dadan, Claire Mail, Notif, Resend Inboxes, Playad, Causo, heal.dev, Whiteboard, DeepCell, Archivist, Mercury 3, and Radaric — comes from launch-day descriptions, so hold all of it as claims until you've kicked the tires yourself.

Mia: If one of these solves a problem you actually have, that's the real test — not a launch post, and not our take on it. Thanks for listening, and we'll be back tomorrow with whatever ships next.