
1118 | Agents Take the Wheel: From Meetings to Money to Your Mac
节目备注
This episode tours the new wave of AI agents: meeting bots that talk on your behalf, agents that run marketing and fix production bugs, agents that move money and build trust, and the MCP-based infrastructure connecting them. We close with the Mac and iPhone tools developers and creators are using right now, from parametric CAD to gesture apps and RAW photography.
Timeline
- 00:00:04 Opening
- 00:00:50 Agents join your meetings and calls
- 00:04:10 Agents running real business operations
- 00:07:40 Agents handling money, deals, and judgment
- 00:10:59 Getting seen in the AI-search era
- 00:13:53 MCP: the plumbing connecting agents to work
- 00:18:55 Mac and iPhone tools for makers and developers
- 00:24:14 Closing
Related links
- Agents API by Recall.ai
- Pinecall
- superlog
- Creatify Max
- Ridian
- Versine
- GatherHQ Agent Research
- TheBold
- AI Visibility by TruLata
- RankSpot
- Clamor
- Link.bio
- Glass Page
- Superhands
- Skiss
- pushr
- AWS Emulator
- Lasso
- 3BCAD
- GlissTouch
- Continuo
- Lampa Camera
- VINC
- Bezzle
- seshka
- Hako
- ExtraDock 5
This episode is produced by Bri. Bri uses advanced AI technology to turn the feeds you care about into podcasts made for listening. Contact us at hi@bri.so.
文字稿
Mia: Welcome back to the briefing, everyone. I'm Mia.
Milo: And I'm Milo. And today we're doing a full sweep of a day of launches — one daily briefing, drawn entirely from the products that shipped in the last twenty-four hours.
Mia: And the thread running through basically everything today is agents. AI agents showing up in your meetings, running your marketing campaigns, handling your money, even getting your company mentioned when someone asks ChatGPT for a recommendation.
Milo: Right, and then there's a second layer to it — this thing called MCP that's becoming the plumbing that connects agents to actual work. Plus, because it's not all robots, a bunch of really sharp Mac and iPhone tools for people who still like building things with their own hands.
Mia: So let's start where a lot of people actually experience this stuff — meetings and phone calls. Because that's changing fast.
Milo: Yeah. So the headline here is Recall.ai. They've launched an Agents API, and the pitch is simple: one API call, and an AI agent joins your Zoom, Google Meet, Teams, or Webex call.
Mia: And I want to be clear about what "joins" means here, because it's more than the transcription bots people are used to. This agent can hear what's said. It can speak. It can show camera video. It can screen-share. It can chat in the meeting. So it's a participant, not a notetaker.
Milo: That's the real shift. For a couple of years, the pattern has been: a bot sits in the corner of the call, silently recording, and you get a transcript afterward. Useful, but passive. What Recall is enabling is an agent that's actually in the room, doing things in real time.
Mia: Now, who's this for? Developers, mostly. It's an API, so you're building an agent that needs to attend meetings — think of anything where a meeting is where the work happens. A scheduling agent, a support agent, a coaching agent. Whatever it is, the hard part — joining the call across four different video platforms and being able to see, hear, speak, and share — is handled by the API.
Milo: And the practical question I'd want answered before trusting it: how does it hold up on a messy real call? Bad audio, overlapping speakers, someone's laptop fan roaring. The launch tells us what it can do, not how reliably it does it under real conditions. That's an open question worth holding onto.
Mia: Related, but a different slice of the same world: Pinecall. This is an open-source voice AI agent platform, built on LiveKit, and it's under the Apache-2.0 license.
Milo: Two things make it interesting. First, testing. You can simulate callers before you put your voice agent in front of real humans, and there's call judging built in — so you're not just hoping your phone agent works, you're actually grading it.
Mia: Second, self-hosting. You can run it with open models, on your own infrastructure. So if you're a company that can't ship customer phone calls through someone else's cloud, that's a genuine option.
Milo: So put the two together and you get the bigger picture: meetings and phone calls are becoming places where agents act, not just places where agents transcribe afterward. That's the why-it-matters line for this whole area.
Mia: And what comes next, if this holds? If agents can be full participants in calls, the obvious question is: what work do they do in those calls? Not just talking — closing tickets, negotiating, scheduling. That's the natural follow-on.
Milo: Which, honestly, is a perfect bridge, because the next set of products is exactly that: agents doing actual operational work, not just conversation.
Mia: Let's take the engineering one first. Superlog — this is a YC company — scans your logs for production bugs that your monitoring missed. Then it investigates around the clock and posts root-cause findings to Slack.
Milo: That last part is what makes it operational rather than just another dashboard. It's not "here's a graph, go look." It investigates, it finds a root cause, and it delivers the finding where engineers already live — Slack. And the claim attached to it is that 87% of fixes from its findings get merged.
Mia: I'd flag that number as a maker claim, not a verified result — it's their figure about their own product. But as a claim, it's meaningful: it suggests the findings are actionable, not noise, because engineers are apparently turning them into merged fixes.
Milo: Who's it for? Teams with real production systems, where the pain is that your monitoring is tuned for known failure modes and the weird bugs slip through. Superlog is aimed at exactly that gap — the bugs your alerting never fires on.
Mia: Now jump to marketing. Creatify Max is positioned as an AI ad teammate. Give it a URL — presumably your product page — and it makes video ads from that. Then it launches Meta and TikTok campaigns and monitors them 24/7.
Milo: And there's a crucial guardrail in the description: nothing goes live without approval. So it's autonomous in the doing — creating, launching, monitoring — but there's a human gate on the actual publish step.
Mia: That approval gate is going to come up again later in this episode, because it's the recurring design answer to "how do we trust an agent with our money and our brand?" Here it's explicit: the agent does all the legwork, a human signs off before anything goes live.
Milo: The unknown here is scale. One approval before one campaign is easy. What happens when an agent is monitoring twenty campaigns around the clock and pinging you constantly? Do the gates hold up, or do humans start rubber-stamping? That's the open question.
Mia: Third one in this group: Ridian. This is an AI-first scheduling suite, and the scope is wide — booking pages, payments, waitlists, funnels, CRM, calendar sync.
Milo: And the agent piece: it supports agents like Claude, ChatGPT, and something called Alvin, and those agents get their own mailbox. So an agent isn't just filling in forms — it's receiving email, presumably handling scheduling conversations end to end.
Mia: So across these three — Superlog in engineering, Creatify Max in marketing, Ridian in sales and scheduling — you're seeing the same pattern: agents moving out of the chat window and into autonomous operations. Not answering questions. Doing jobs.
Milo: And the honest summary of where we are: the shift is real and the pieces are shipping, but the trust infrastructure — approval gates, monitoring, whatever liability framework eventually applies — is young. Which is exactly where the next group of products lives.
Mia: Money first. There's a product called Versine, described as "Venmo for AI agents."
Milo: Which sounds like a joke until you realize someone had to build it, because the moment agents do real work, they need to move real money. Here's how it works: agents send and request money via MCP — that Model Context Protocol we'll come back to — and there are two safety models. Passkey approval, where a human authorizes each transaction, or trusted-contact auto-send with limits, where pre-approved transfers can happen automatically up to a cap.
Mia: And one detail that's actually clever: no account is needed to receive. So if an agent needs to pay someone — a contractor, a vendor — that person doesn't have to sign up for anything. Lowering that friction matters a lot for agent-to-human payments.
Milo: Now the flip side of delegation: research and judgment. GatherHQ has something called Agent Research, which runs panels of agents that answer surveys. And the design detail I like: every answer is labeled as evidence, inference, or unknown. So you're not getting a confident-sounding mush — you're getting the epistemic status spelled out.
Mia: There's user-controlled permissions, it's in beta, unpaid, and they say no PII is involved. So it's early and deliberately constrained. But the idea — using agent panels for research, with honesty labels on each claim — is an interesting answer to the "can I trust what an agent tells me" problem.
Milo: And then there's fundraising. TheBold turns founder execution into investor meetings. The flow: you show investor criteria, a verified track record, AI simulations — and then you book calls directly.
Mia: The word "verified" is doing real work there. In fundraising, the bottleneck is often trust — does the investor believe you did what you say you did? If TheBold can put a verified track record in front of the matching step, that changes the top of the funnel from a cold email to a qualified, evidence-backed introduction.
Milo: Worth being careful, though: "AI simulations" of investor conversations is a claim about capability, and we don't know how predictive those simulations actually are. Simulating a meeting is easy; simulating an investor's real judgment is hard.
Mia: So the theme across Versine, GatherHQ, and TheBold: money, hiring, and research decisions are being delegated to agents. And the open question hanging over all three is liability. If an agent overspends, or overclaims in a fundraising simulation, or an agent panel gives you a confident wrong answer — who's responsible?
Milo: Nobody in this batch answers that. It's genuinely unsettled. And it connects to something adjacent — because if agents are making decisions about what to buy and who to trust, there's a whole other market forming around getting agents to notice you in the first place.
Mia: Here's the framing: search used to be about ranking in a list of blue links. Now, increasingly, buyers ask ChatGPT or Gemini or Claude a question, and the AI answer names two or three businesses. If you're not one of them, you don't exist in that conversation.
Milo: Two products attack this from different angles. The first is TruLata, with AI Visibility. Free check: does ChatGPT, Gemini, or Claude name your business when buyers ask? And then weekly tracking across nine AI engines.
Mia: The free check is a smart wedge — every business owner can ask themselves "am I in the answer?" and now there's a tool to actually test it. The weekly tracking across nine engines is the ongoing product. The obvious unknown: these engines update their weighting and their knowledge all the time, so visibility can swing for reasons outside your control.
Milo: The second is RankSpot, and it's more aggressive. It's an SEO and AI-search execution layer for agents, via MCP. It writes and publishes articles at $3 per article. It runs prompt checks at $0.004 each. It pitches backlinks. No subscription — you pay per unit.
Mia: Per-unit pricing is notable because it makes it agent-friendly. An agent with a budget can decide "we need ten articles and fifty prompt checks this month" and just execute, without a human negotiating a SaaS contract. That's a pricing model built for the agentic era.
Milo: Two supporting products round out this picture. Clamor is a social intelligence platform — it turns social signals into actionable insights, analyzes public accounts without requiring a login, has an API, and an "Ask Clamor" chat. So if AI answers draw on what's being said publicly, understanding those signals is part of visibility.
Mia: And Link.bio is on the personal side. It's a link-in-bio builder where you paste your profile and AI creates the page — and via MCP, agents can manage it. There's a free lifetime Premium code, PHLINKBIO, in week one. Small product, but same logic: your presence now has to be legible to both humans and agents.
Milo: So the open question for this whole area: the engines that decide who gets cited are black boxes, and their weighting changes over time. Anyone building a business on being cited by ChatGPT is building on ground that can shift. Worth tracking, not worth panicking about yet.
Mia: Which brings us to the thing that's been lurking in the background this whole episode: MCP. The Model Context Protocol. It's showing up in RankSpot, in Versine, in Link.bio — and it's emerging as the standard connective tissue between agents and tools.
Milo: Let's define it plainly for listeners: it's a protocol that lets an AI agent connect to external systems in a standard way. Instead of every integration being bespoke, a tool exposes an MCP server, and agents that speak MCP can use it. That's why you keep hearing "via MCP" attached to everything today.
Mia: Now, the products building on it. Glass Page is the context one. Agents publish their work as things called Panes. Teams comment in Threads. The agent replies with new Versions. And there's one Latest Link that's always current.
Milo: That's a collaboration loop designed for humans and agents working together. The agent produces something, humans critique it, the agent revises — and nobody has to wonder which version is current, because the Latest Link is always the live one. It solves the "which draft is real?" problem that plagues any review process.
Mia: Superhands is the rules one. It turns designs into guidelines that agents follow, via MCP, inside Claude, ChatGPT, Cursor, and Lovable. The pitch: fix the system once, instead of correcting the agent every session. Free tier is 100 credits a month, Pro is $29.
Milo: Every developer who's used an AI coder knows this pain. You explain your conventions, the agent half-follows them, and next session you explain again. Superhands is trying to make your standards a persistent thing the agent actually loads. That's a real problem with a concrete answer.
Mia: Skiss is the reference material one. It saves sites with full-page captures, plus fonts, colors, and sections extracted from the real CSS — not guessed, pulled from the actual stylesheet. It's a searchable library, and it exposes something called Design.md via MCP for Claude, Cursor, and Figma. Fifty sites free.
Milo: So when your coding agent is building a UI, instead of a vague "make it look modern," it can read the extracted design tokens from a real site you saved. Real CSS means real values — actual hex codes, actual typefaces. That's much better grounding than a description.
Mia: And Pushr is the reliable channel one. It sends iPhone push alerts from any script or any AI agent via one HTTP request. Live Activities support, must-ack alarms — alarms that demand acknowledgment — and uptime checks. MIT licensed and self-hostable.
Milo: Why does an agent need push alerts? Because an agent that monitors something around the clock — remember Superlog and Creatify Max monitoring 24/7 — needs to reach a human when something matters. A must-ack alarm is the serious version of that: not a notification you can swipe away, one that keeps demanding acknowledgment.
Mia: There are two supporting products here too. An AWS Emulator — free, local, lets you build and test AWS apps using S3, DynamoDB, SQS, without deploying or even having an AWS account. And Lasso, a Mac app where you draw arrows and notes directly on your screen with a shortcut, screenshot marked-up areas, and — the MCP part — a local MCP server that lets Claude or Cursor actually see your annotations. Nine ninety-nine, one time.
Milo: That Lasso detail is quietly great. Instead of writing "the button is too close to the header" in a comment, you draw on the thing, and your coding agent sees the drawing. Visual feedback into the agent loop.
Mia: So the why-it-matters for this whole section: agents need three things to be genuinely useful — shared context so they know the situation, rules so they don't violate your standards, and reliable channels to reach you. This batch covers all three.
Milo: The big unknown: will one MCP ecosystem dominate, or will it fragment? Right now it looks like a standard, but standards wars have happened before. For now, building on it looks like the reasonable bet.
Mia: And finally — the ground this whole wave stands on: the tools humans use every day. Because while agents automate, the local tools for hands-on creators keep getting sharper.
Milo: Start with 3BCAD. Free, open-source, written in Rust, parametric CAD — so you define your model with parameters and it recomputes — with built-in 3D-print prep, assemblies, and 2D drawings. Apache-2.0 license, cross-platform, and it imports STEP files.
Mia: The significance is the whole package in one free tool. Historically you'd stitch together a CAD package, a slicer for print prep, and a drawing tool, or pay for a commercial suite. Having parametric modeling, assemblies, drawings, and print prep natively — and STEP import so you can bring in models from other tools — makes this a serious option for hobbyists and small shops.
Milo: On the input side of the Mac: GlissTouch. A menu bar app that adds custom trackpad gestures — multi-finger taps, shapes, edge slides — bound to over 95 actions. Free for up to 5 gestures.
Mia: "Shapes" is the part I find interesting. Drawing a shape on the trackpad to trigger an action is a gesture vocabulary most trackpad tools don't have. If you're a keyboard-averse person, that's a lot of reachable functionality.
Milo: For multi-Mac people: Continuo. One keyboard, mouse, and trackpad shared across your Macs over LAN. No Apple ID involved — pairing is just three words. There's a Screen view with drag-and-drop between machines. Free for 2 Macs, Pro is $39, one-time payment.
Mia: The no-Apple-ID, three-word-pairing detail matters because the existing solution for this requires the same iCloud account on every machine. If you have a work Mac and a personal Mac, that doesn't work. Continuo sidesteps that entirely, and one-time pricing instead of a subscription is maker-friendly.
Milo: And for the photographers: Lampa Camera on iPhone. Custom Bayer RAW processing — that's working with the raw sensor data directly — with film-inspired profiles, and deliberately no HDR, no AI, no sharpening.
Mia: That's a philosophical position as much as a feature list. Almost every modern camera app processes aggressively. Lampa is saying: here's the sensor's actual data, interpreted through film aesthetics, without computational make-up. 450,000 downloads and a 4.7 rating suggest an audience for that — though those are reported figures, not independent verification.
Milo: Rapid-fire through the supporting Mac and iPhone items, because they're all small but real. There's an AI filmmaking studio called VINC — one browser environment that goes script to storyboard, does 3D shot blocking, taps 39 generation models, and includes an editor with sound effects and background music.
Mia: Bezzle is a free full design studio for Mac and Windows that opens Figma, XD, Sketch, and Photoshop files — with components, prototypes, dark mode, and an MCP AI assistant. So it's both a free alternative design tool and an interop play.
Milo: seshka is a free, MIT-licensed menu bar app for a very specific pain: leftover Claude Code processes. When you close a session, the dev servers and ports and memory it spawned can keep running. seshka kills those, grouped by session, and can end idle sessions automatically.
Mia: Hako is a free iPhone and Mac app — no account needed — that lists all your coding agents, like Claude Code and Codex, across all your Macs, sorted by which one needs you. You can answer from the Lock Screen or your Watch. So as you run more agents, you need a triage view of them. That's what Hako is.
Milo: And ExtraDock 5 lets you build unlimited custom docks on every screen and edge — per-dock apps, widgets, profiles. 50% off with the code BLACKFRIDAY50 until December 1.
Mia: So what's the through-line for this last section? While agents automate the operational layer, the local tools for makers keep getting sharper — better CAD, better gestures, better multi-machine setups, more honest cameras. These are tools for the people doing the building.
Milo: The open question is where agentic features show up in these tools next. Bezzle already has an MCP assistant. When does your CAD tool have an agent that checks your design for printability? That boundary is going to keep moving.
Mia: That's the briefing. Agents in your meetings, agents running operations, agents handling money and judgment, the new fight to be seen in AI answers, the MCP plumbing underneath it all, and the hand-tools that haven't been automated yet.
Milo: Everything today was maker-described, so treat the numbers and claims as claims — the promising parts are real, and so are the open questions.
Mia: Thanks for listening — we'll see you at the next briefing.