1015 | Agents, Models & Everyday AI

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

Milo: And I'm Milo. Today we've got a packed briefing—lots of launches across benchmarks, agents, memory systems, hardware, and a bunch of everyday tools.

Mia: The thread connecting all of this today is that AI is moving out of the chat window and into the plumbing—models that are cheaper, agents that do actual work, and the infrastructure that keeps it all standing up. So let's get into it.

Milo: Let's start at the foundation, actually—the models themselves. Because a couple of benchmarks dropped that suggest the gap between the big frontier models and the smaller, cheaper ones is narrowing.

Mia: Right. So Nimble-v4 and v4-max, their decision models, beat Jev 1.13 on the Decision Index. And the v4-max version hit the top score, 68.46. What's notable there is the model size—it's a 27B model. That's small. That's not a giant frontier model.

Milo: And that's the story, isn't it? A 27 billion parameter model topping a decision benchmark. If that holds up, it means you can get frontier-adjacent decision quality at a fraction of the cost and hardware requirements.

Mia: Same pattern on the coding side, with a product actually. It's called Reason—it's a coding agent that runs cloud or local, and it has an optimized cache, which is how they keep costs down.

Milo: And the claim is it scored 65.1 versus Codex's 62.5 on the Coding Agent Index, and it's 22 percent cheaper. Now, again—these are the makers' own benchmark claims. We should treat them as claims, not independently verified results. But if they're even roughly accurate, the picture is consistent: smaller and cheaper options closing in on the frontier.

Mia: And why does that matter for everything else we're going to talk about today? Because if the underlying model cost keeps dropping, all the agent products built on top become more affordable to run—and there's a whole wave of agent tools in this briefing that only make economic sense if inference is cheap.

Milo: One thing to flag as an open question: we don't know whether these gains hold across all task types. Benchmarks are narrow slices. A model that tops a decision index or a coding index might not generalize. So, worth watching, but with some skepticism.

Mia: Which brings us nicely to those agent tools. And the first one is aimed at something very concrete—getting leads, doing outreach, even hunting for jobs. It's called Retriever AI.

Milo: So this is a browser agent, and it's free, and the makers claim 81.4 percent on Web Bench, five times faster than rivals, and three hundred times cheaper. Again—claims from the makers. But the pitch is clear: instead of you clicking around the web filling in forms and scraping contact pages, an agent does it.

Mia: Who is this for? Recruiters, salespeople, job seekers—anyone whose work involves a lot of repetitive web browsing to find people or opportunities. If the benchmark numbers translate to real-world reliability, that's a meaningful productivity change. That "if" is doing a lot of work though. Browser agents have a history of being demos that impress and then fail on messy real pages.

Milo: Now, a different flavor of agent—this one sits on your desktop. It's called AirJelly, and it's free on Mac and Windows. It watches your screen context, automatically creates tasks based on what it sees, and sends reminders.

Mia: So imagine you're working and an email comes in about a deadline—AirJelly picks up on that context and turns it into a task with a reminder. The privacy question is obvious here: something is watching your screen all day. We don't have details on how that data is handled, so that's a real open question. But the idea of an agent that's proactive rather than waiting for you to type a prompt is a notable shift.

Milo: And then there's the marketing vertical, which is getting its own agent stack. AdAnt 2.0 pitches itself as a 24/7 AI marketing team, with four agents: a Strategist, a Creator, an Analyst, and a Manager, all working on your social presence. It comes with 40 free credits to try.

Mia: So instead of hiring or being the marketing team yourself, you're delegating to four specialized agents that hand off to each other. That's a real shift in how a small business could operate. But again, "24/7 AI marketing team" is the maker's description. What we don't know is the quality of what those agents actually produce and whether the Analyst agent's insights are any good.

Milo: And there's a fourth one in this space that's more infrastructure-flavored. It's called Pancake, and it describes itself as "Vercel for GTM"—go-to-market. It was rebuilt around MCP, the Model Context Protocol, and it gives agents what they call a GTM Brain, access to 40-plus data providers, and sending infrastructure.

Mia: So the analogy is: Vercel made deploying apps trivially easy; Pancake wants to do that for go-to-market operations, but for AI agents rather than human developers. They say it hit number one in May and has since been rebuilt. Interesting sequencing—rebuilding after traction suggests they learned something about what agents actually need.

Milo: What connects all four of these is that agents are moving from answering questions to doing jobs—sales, productivity, marketing. And the honest caveat across the board is the same: benchmark and demo performance doesn't tell us how reliable any of these are on a random Tuesday with a messy real-world task.

Mia: Now, if agents are going to do real work over days and weeks, they need something most of today's tools lack: memory. And that's the next thread—memory systems and the protocols that connect them.

Milo: The standout here is Hindsight 1.0. It's open source under an MIT license, and the makers claim it's number one on the Agent Memory Benchmark. Under the hood there's a Rust engine called Memlake, and the claim is it handles 100 million memories with under 100 milliseconds at p99 latency.

Mia: Those are the kinds of numbers that matter, because agent memory is only useful if it's fast at recall. If your agent takes seconds to remember something, it's not practical. And being open source with MIT licensing means anyone can adopt it without lock-in. Again—benchmark claims from the makers, but the architecture story is at least concrete: Rust, 100M memories, sub-100ms p99.

Milo: Then there's Make0 Brain, which attacks a different part of the memory problem: portability. It stores memory across ChatGPT, Claude, and Cursor via MCP, and you can import your existing conversations in about two minutes. It's in open beta and free.

Mia: So the problem it solves is that your context is trapped in silos right now. What you told ChatGPT last week, Claude doesn't know. Make0 Brain is trying to make that memory travel with you across tools. That's genuinely useful if it works as described, because re-explaining yourself to every new assistant is one of the most annoying parts of using these tools daily.

Milo: And daily.dev, which many developers know as a news feed, is getting in on this too. daily.dev for agents opens MCP access for Claude, Cursor, and Codex to their feeds, search, and bookmarks, plus they're launching a plugin marketplace and an API.

Mia: So your agent can read the same developer content you do. It's another piece of the same puzzle—giving agents access to context they currently don't have. And notice MCP keeps coming up—Make0 uses it, daily.dev uses it, and we'll see it again in a moment. It's becoming the connective tissue, and the thing to watch is whether standardization consolidates around it.

Milo: Which is a good segue, because the next group is coding agents—not just writing code, but shipping and maintaining it. First up: Exodus. It's a free, open-source tool where coding agents migrate apps to Cloud Run, using more than 50 playbooks, and it gives you cost reports.

Mia: So the tedious part of deploying—figuring out the migration, the configuration—is handed to an agent with a playbook library, and you get visibility into what it costs. That's practical: the playbooks encode institutional knowledge about how to do the migration properly, rather than letting an agent improvise.

Milo: Similar territory, different angle: WriterzRoom is a multi-agent platform with eight LangGraph agents and something they call a Quality Gate. It runs 100 percent serverless on Cloud Run across nine services.

Mia: So this is agents building software with a quality checkpoint built in. The Quality Gate idea is important—output doesn't just flow through; something evaluates it. That's the maturing of the pattern: not "AI writes code," but "AI writes code, gets checked, then it ships."

Milo: And on the maintenance side, LegacyFixer—this one's from GitHub—scans Python dependencies for vulnerabilities and opens pull requests, but only if its own validations pass.

Mia: That "only if validations pass" part is the whole design. Automated dependency upgrades are risky because they can break things silently. A tool that refuses to open a PR unless its checks pass is being appropriately cautious. This is the unglamorous but essential work of keeping software safe over time.

Milo: And then there's Tiden, which tackles the trust question from another direction. It does intent verification for coding agents—versioned specs, and verdicts from real tests. And notably, no AI judging the results.

Mia: I want to underline that. The failure mode everyone worries about with AI-written code is an AI grading its own homework—rubber-stamping its own output. Tiden's pitch is that the verdict comes from real tests, not from another model's opinion. That's a meaningful design choice, and it addresses a genuine gap.

Milo: So the overall shift here: from agents that write code to agents that ship it, verify it, and maintain it. And the natural next question is—what about the code that's already being built carelessly? Which brings us to securing AI-built apps.

Mia: Epherem is a free security scanner specifically for AI-built apps. You can scan a ZIP or a GitHub repo, and it generates fix prompts you can feed straight into Claude Code, Cursor, or Lovable.

Milo: That last part is the clever bit—it doesn't just tell you there's a problem; it produces the prompt that gets your coding agent to fix it. So the security review becomes part of the AI build loop rather than a separate step a human has to translate.

Mia: And on the steering side—how humans direct these agents—there's EditUI. It's a free Chrome extension, no signup, that lets you point at UI elements on a page and send batch instructions to Claude Code or Cursor.

Milo: So instead of writing "change the button in the header, move the form, adjust the spacing" in text, you literally point at the elements and describe what you want. That's a much more direct and visual interface between humans and coding agents. It removes a whole layer of ambiguity in describing UI changes.

Mia: These two together close the developer loop: build with agents, secure what they built, steer them more precisely. Trust in AI-written code is becoming an actual product category.

Milo: And under all of this is infrastructure that has to actually work. Let's talk about the reliability layer, because there are three tools here that all attack the same problem from different angles.

Mia: First, Inferock. It's an LLM inference proxy, and its distinguishing feature is receipts for failed calls—so when an inference call fails, you get documentation of it, and they credit back on verified failures. Their Watch product is free if you use your own API keys.

Milo: That's interesting because inference failures are usually invisible. You pay, the call silently fails or degrades, and you have no recourse. Inferock is essentially adding accountability to a layer where nobody expected it. Whether a proxy can sustain that business model is an open question, but the transparency angle is genuinely new.

Mia: Second, Specreel. It generates demos from Playwright traces or from crawling a URL, and then republishes them on green builds—so your demos never go stale.

Milo: Every developer knows the pain of stale demos. The screenshot on your landing page shows version one; you're on version twelve. Specreel's pitch is that demos regenerate automatically when builds pass. Small idea, but it solves a real, persistent annoyance.

Mia: And third, Cloud Releases—a single feed of updates across AWS, Azure, GCP, and Cloudflare, covering over 700 services, with deprecations highlighted, and it's free.

Milo: If you operate anything on cloud infrastructure, keeping up with changes across four providers is its own job. A unified feed with deprecations surfaced—because those are the ones that break things—is the kind of tool that quietly saves outages.

Mia: The theme across all three: reliability and visibility are becoming products in themselves. As AI systems stack on top of fragile infrastructure, someone has to sell the guarantee that things work.

Milo: Let's shift to hardware now—the physical layer. Two stories, one on your desk and one in the datacenter. Starting with the home: HomePod mini 2.

Mia: It's 149 dollars, it has something called Sound Glow, and it has better integrations with Mac and Apple TV. But here's the notable gap: the AI version of Siri is not there at launch. It's promised in a future update.

Milo: And that's a real tension. You're buying AI hardware, essentially, and the headline AI feature isn't ready. The hardware is being positioned ahead of the software, which tells you something about where the industry is—everything is racing toward AI, and sometimes the race outruns the actual capability.

Mia: Now the datacenter side: Microsoft announced the Ma 300, a next-generation AI inference chip targeting 2027 production. The claim is 40 percent more tokens per megawatt, and it deploys in racks of 96 accelerators.

Milo: Tokens per megawatt is exactly the right metric to watch, because at datacenter scale, power is the constraint, not chips. If that 40 percent claim survives real-world deployment—big if, since it's 2027 and unproven—it directly lowers the cost of everything we've been talking about: cheaper inference, cheaper agents.

Mia: So hardware from your desk speaker to the rack is being reshaped around AI, but the Siri delay is a reminder that the software isn't keeping pace with the hardware ambition.

Milo: And on that note about voice—let's get into the fun part of the briefing: voice and vision tools in daily life. There are five of these and they're a diverse bunch.

Mia: Starting with the most unusual one: Winq. It's macOS hands-free control via your camera. You focus with your gaze, you click by winking—with each eye doing something separate, apparently—and there are eleven gestures. And it's all on-device for privacy.

Milo: Per-eye wink clicking is a bold interaction design. And the on-device processing is the key detail—camera data controlling your computer is only acceptable if it never leaves the machine. For accessibility especially, this could be genuinely important. Whether gaze-and-wink feels natural for everyday use is the big unknown.

Mia: Then on the voice synthesis side: Kugel 3.5 TTS. The makers claim it beats four rivals on eight out of nine tests. Two details stand out: it gets long numbers right 97.1 percent of the time—which is a real problem area for TTS, reading things like phone numbers and prices—and it supports IPA, the phonetic alphabet, so you can specify exact pronunciations. It costs 5.25 cents per minute.

Milo: The long-numbers stat is the one that tells you they're solving actual problems, not just making pretty voices. Numbers are where TTS systems embarrass themselves constantly.

Mia: Then there's Mira Tours, an AI walking audio tour guide app. You can build a tour by chatting with it, or it generates one automatically, it works in 23 languages, and you get 30 free minutes with pay-per-use after.

Milo: So instead of a pre-recorded tour, you get one generated for wherever you are, in your language, and interactive. Travel is a natural fit for this kind of AI.

Mia: Table Canon is more niche but really interesting if you play tabletop RPGs. It turns your game-night audio into session recaps, dossiers of the entities in your campaign, and a searchable history. Six hours free.

Milo: If you've ever run or played in a long campaign, you know the pain: three months in, nobody remembers what that NPC promised you in session two. A searchable campaign history is solving a real, felt problem for that community.

Mia: And the last one in this group is for hiring: Screenwell, with an AI interviewer called Luna that runs first-round screenings on Meet or Teams, producing ranked scorecards. It works in over 30 languages, 14 days free.

Milo: First-round screening is high-volume and repetitive, so it's a logical place for automation. But AI in hiring draws justified scrutiny—consistency, fairness, and what a "ranked scorecard" actually measures are all open questions. That one will need more evidence before we'd call it solved.

Mia: Across this whole group, the pattern is voice and vision AI moving into personal, social, and hiring workflows. And what decides adoption is exactly the two things we keep flagging: privacy—does the data stay on-device, like Winq—and accuracy, like Kugel's numbers.

Milo: Alright, last thread: everyday software. Experimentation, money, workflows—products that look conventional but are being reshaped by this era. First up, Confidence by Spotify.

Mia: Yes, Spotify of all companies is shipping an experimentation platform. It's warehouse-native—which means it works on top of your existing data warehouse—with A/B testing, feature flags, and session replay. It's free to start.

Milo: That's a notable move. Spotify runs experiments at enormous scale internally, and packaging that as a product, warehouse-native so your data doesn't have to move, is a direct challenge to the established experimentation vendors.

Mia: Then something completely different: NoteHouse. It's software for private loans—it creates promissory notes, handles e-signatures, and keeps shared accounting between the parties.

Milo: Private loans between friends or family are usually handled with a text message and a handshake, and that's how relationships get ruined. Formalizing that with proper documents and shared books is unglamorous and genuinely useful.

Mia: Rambox also got a redesign—it's an app that aggregates your messaging and work tools—and the update brings a Command Center, Unified Search, internal tabs, and workspaces for separating accounts and contexts.

Milo: So if you juggle five Slack workspaces and three email accounts, workspaces let you keep those contexts separate instead of one chaotic sidebar. It's workflow hygiene, but for a lot of people that's daily quality of life.

Mia: And the last one is Blade, and this one comes with a privacy conversation attached. It mines contacts from your email—but importantly, only from headers, not body content—to help you find contacts for product launches. It was cofounded by the cofounder of KAYAK, and it's free for 30 days.

Milo: The header-only detail is their answer to the obvious privacy concern—they're claiming they never touch the content of your emails, just who you correspond with. And the KAYAK pedigree gives it some credibility. But it's still a tool whose entire premise is reading into your mailbox, and that tradeoff is exactly the kind of thing people should scrutinize.

Milo: It's a pattern across this whole group, actually: AI-era capabilities wrapped in familiar, everyday products, and the data-access questions never fully go away.

Mia: So let's pull it together. What's the through-line of this whole briefing?

Milo: It's cost and autonomy coming down simultaneously. Cheaper models—the 27B topping the Decision Index, Reason undercutting Codex—make agents economically viable. Agents then move into real work: sales, marketing, coding, hiring. That creates demand for memory that persists, protocols like MCP that connect tools, and infrastructure—security, verification, reliability—that keeps it all trustworthy.

Mia: And the honest caveats to carry with you: these benchmark numbers are maker claims, not verified results. We don't know how they hold up across task types or in messy real-world conditions. The HomePod shows hardware shipping before its AI software is ready. And tools like Blade remind us that convenience and data access are always trading off.

Milo: That's the briefing. Thanks for listening—talk to you next time.

Mia: See you then.