0924 | AI Agents on the Rise: The Week's Tools

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

Milo: And I'm Milo. Today we've got a full slate of launches to walk through, and there's a thread that ties almost all of them together: AI that doesn't just answer questions, but actually does work. Agents with memory, agents with budgets, agents that publish content, agents that need to be scored and secured like employees.

Mia: Right, and it's not just the agent products themselves. It's the infrastructure around them, the models underneath, and then, interestingly, a bunch of tools at the end that deliberately push back against all of it. Tools with no AI at all. So we'll move from autonomy to accountability to memory to the models, and then land on the human-centered side of things.

Milo: Let's start with the most vivid version of the agent idea. Pactto. It's built for creative teams, and what it does is give you persistent AI rooms.

Mia: Persistent rooms. So the idea is that instead of starting a fresh chat every time, the room keeps its context. Everything the team has reviewed, everything that's been said, stays there.

Milo: Exactly. And the pitch on the review side is video review with what they call studio quality. So a team working on video can bring their work into the room, get feedback, and here's the part that makes it different from a chatbot: the agents apply that feedback in live.

Mia: That's the jump I want people to notice. A normal AI tool would say, "here's my suggestion for your edit." Pactto's claim is the agent actually makes the change. It reads the room's context, the feedback that's accumulated, and it acts on it.

Milo: Who's that for? Creative teams, clearly. Post-production, editing, anyone who does review cycles on video. And anyone who's been through those cycles knows the pain. You leave timestamped comments, someone has to interpret them, translate them into edits, go back and forth. If the agent can genuinely take "fix the pacing in the second act" or specific feedback and apply it while keeping all the room's history, that removes a whole layer of manual work.

Mia: But here's the honest limitation, and we should flag it with anything in this category: we don't know how reliable this is in the real world. "Applies feedback live" is a claim from the maker. Whether the agent gets the feedback right, whether it understands nuance in creative direction — that's unverified. Creative judgment is exactly where AI agents tend to stumble.

Milo: And there's a real open question about how much context matters. The persistence is the selling point, but the value of persistence depends entirely on whether the agent uses that context well over weeks and months, not just in one session.

Mia: Okay, staying in the agent world, but moving from creative teams to marketing. This next one makes some of the boldest claims we've seen. Naise AI.

Milo: Naise runs autonomous marketing agents. Social, influencer work, PR. You set it up and the agents execute. And the pricing is surprisingly accessible — starting at 59 dollars 90 a month.

Mia: Now the claims. They say it saves you 40 hours a week and cuts costs by 80 percent.

Milo: Which is... a lot.

Mia: It is a lot. And we have to be careful here: these are the maker's numbers, not results we've verified. Forty hours a week is essentially a full-time employee. Eighty percent cost reduction on marketing spend. If those numbers were broadly true, every marketing team would be restructuring tomorrow. So treat them as claims.

Mia: What we can say concretely is what the product does: autonomous agents covering social, influencer, and PR execution, at a price point that's far below what an agency or even a junior hire costs.

Milo: And there's a comparison worth drawing here. Compare Naise to something like hiring help. A junior marketer costs thousands a month. Naise is 60 dollars. The question isn't whether it's cheaper — it obviously is — it's whether the output is usable. That's the unverified part.

Mia: And it pairs naturally with another product in the same space: RankControl.

Milo: RankControl is interesting because it's built around seven agents. These agents create and publish content on your own domain — so it's not posting on social, it's publishing to your site.

Mia: And then it tracks something new: citations in six different AI engines.

Milo: Right, and that's a genuinely different angle. As people search less on Google and ask AI assistants more, businesses are starting to care whether ChatGPT or Claude or whoever mentions them when someone asks a question. RankControl writes content, publishes it, and then monitors whether the AI engines pick it up and cite it.

Mia: So it's SEO thinking applied to the AI era. Instead of "are we ranking on page one," it's "does the model know we exist."

Milo: Pricing is simple: one plan, with a 7-day trial. So you can at least test it before committing.

Mia: Between Naise and RankControl, you get two flavors of marketing autonomy — social and PR on one side, content and AI-visibility on the other. Both unproven in terms of results, but both pointing at the same future: marketing as something you supervise rather than something you do.

Milo: Which brings up an obvious problem. If agents are doing the work, how do you know if they're any good? That's where the next two products come in, and honestly, this might be the most important part of today's conversation.

Mia: First: AgentScore, from a company called Latitude. And the idea is beautifully simple. Every day, your production agent gets a score.

Milo: On five dimensions: results, reliability, cost, speed, and safety.

Mia: Think about that for a second. That's basically a performance review for software. Did it accomplish the outcome, could you depend on it, what did it cost, how fast was it, and did it do anything unsafe along the way?

Milo: And it's open source, which matters a lot here. Scoring agents is something everyone is going to want to do, and nobody should have to trust a black box for. If the scoring methodology is open, teams can adapt it, audit it, trust it more.

Mia: The reason I think this matters so much: look back at everything we've discussed. Pactto agents applying feedback. Naise agents running your social. RankControl agents publishing to your site. All of those need exactly this — a daily, structured answer to "is my agent actually performing?"

Milo: Without measurement, you're trusting marketing claims — like Naise's 40 hours a week — forever. With something like AgentScore, you'd be generating your own evidence. Daily. Across five dimensions that actually capture the tradeoffs.

Mia: And notice what the five dimensions include: safety. Which is a perfect bridge to the second half of this topic, because measurement alone isn't enough. You also need guardrails. That's Koreshield.

Milo: Koreshield is security specifically for AI support agents. And the setup there is genuinely dangerous territory: an agent that talks directly to your customers, reads from your knowledge base, and calls tools that can act on real systems.

Mia: Koreshield filters the messages coming in from customers, filters the content pulled from RAG — the retrieval side — and filters the tool calls going out.

Milo: So it's monitoring all three sides of the interaction: input, knowledge, and actions.

Mia: And the rollout strategy is smart. They have a detection-then-enforcement mode. So you start in detection mode, where Koreshield just watches and logs what would have been blocked. You review the logs, you see what the agent was actually doing or being fed. And once you're confident, you switch on enforcement and it actually starts blocking things.

Milo: That's how you'd want to deploy any security layer — observe first, act second. And the logs give you an audit trail, which support teams need anyway.

Mia: Put AgentScore and Koreshield together and you see the pattern: measurement and security are becoming their own layer of the stack. A whole category forming around not building agents, but keeping them honest.

Milo: And if agents are going to be measured and secured like team members, they need one more thing employees have: institutional knowledge. Which brings us to memory and data infrastructure. Two products here, and they solve different halves of the problem.

Mia: First half: raw knowledge. That's Alexandria, from Firecrawl. It's a data library that AI agents can access, through MCP, CLI, or API.

Milo: So instead of your agent being limited to whatever's in its training data or whatever you paste in, it pulls from a structured library of data at runtime.

Mia: And the claim from the maker is a 21 percent improvement in answer quality when agents use it. Again, a claim — but at least it's a specific, testable one.

Milo: And there's a signal here that's worth noting: they raised a 75 million dollar Series B.

Mia: That's real money. Investors betting that agent data infrastructure is a large market.

Milo: Now, Alexandria is about giving agents knowledge. The second half of the problem is shared context — the team's accumulated state. That's GBrain.

Mia: GBrain gives you shared memory in markdown — so it's human-readable, you can actually go read and edit what your agents know — plus connected accounts that all of your AI tools can access.

Milo: That last part is the sleeper feature. Think about how fragmented things are right now. Your chatbot has one set of integrations, your coding assistant another, your writing tool another. GBrain's pitch is a single layer of memory and connected accounts that any AI can use. Your context stops being trapped inside each tool.

Mia: Pricing: a team workspace at 99 dollars for the first month, then 199 after that. So it's priced for real teams, not individuals.

Milo: And there's a third, smaller piece in this space worth mentioning, because it connects: Jev State. It's free and open source, and it does something a bit different — it takes your conversations with AI and turns them into executable tests and TypeScript code.

Mia: I like that one because it captures the ephemeral stuff. So much of what happens with AI agents lives in chat logs and disappears. Jev State says: that conversation where you and the AI figured out a behavior? Turn it into a test. Export it as JSON or a SKILL.md file. Make the knowledge durable and verifiable.

Milo: So you've got Alexandria for external knowledge, GBrain for shared team memory, and Jev State for capturing what emerged from your own AI sessions. Three layers of the context stack.

Mia: All of this infrastructure — memory, data, scoring, security — sits on top of the models. So let's talk about the model news, because there's one item: Claude Opus 5.5.

Milo: Anthropic's first model in the 5.5 family. The focus is agentic coding and knowledge — which is exactly the workload everything we've talked about depends on.

Mia: And the headline economic change: it's 40 percent cheaper than Opus 5.

Milo: That's significant. Agent workloads are expensive because agents do many steps — they read, they call tools, they retry. Cost per token matters more for agents than for single chat messages. A 40 percent price cut on a frontier model for agentic coding changes the economics of running agents continuously.

Mia: It's also worth reading the focus. "Agentic coding and knowledge" is Anthropic saying out loud that the primary workload is agents, not chat. The model is being shaped for the tools we've been discussing all episode.

Milo: Okay. So we've spent a lot of time on autonomy — agents that work like teammates, run your marketing, get scored and secured, and remember things. Now I want to pivot, because there's a whole other cluster of tools that go the opposite direction. They put AI in service of the human, moment by moment, and the human stays in control. And the clearest example of that shift is in real-time voice translation.

Mia: Speechka. And the pitch here is wild if it works as described: real-time voice translation in your own voice.

Milo: Not a generic synthesized voice — yours. You speak, and a moment later your words come out in another language, sounding like you.

Mia: The specs: 44 languages, and latency around one and a half seconds.

Milo: One and a half seconds is the make-or-break number. At that latency, this isn't turn-taking translation where you finish, wait, and someone answers. It's close to conversational flow. You could actually have a dialogue.

Mia: Availability is broad too: macOS and Windows apps, plus a free browser version. So you can try it in the browser without installing anything.

Milo: Think about the use cases: international calls, meetings, presentations to multilingual audiences. Preserving your own voice matters more than people expect — tone, emphasis, personality all come through.

Mia: Now compare that with a completely different take on translation: Linguo Translate. This one is a macOS menu bar translator.

Milo: So it lives in the top of your screen, always one click away.

Mia: And the interesting technical detail: 22 languages offline, using Apple's Translation technology.

Milo: Offline is the key word there. No internet needed, which also means your text doesn't leave your machine — a real privacy angle.

Mia: And there's optional AI on top, if you want it for cases where the offline translation isn't enough. Price: 3 euros 99. One-time, presumably.

Milo: So you've got two very different philosophies. Speechka is about live human conversation, many languages, your own voice. Linguo is about quick text translation at your fingertips, private and offline, cheap. Neither is wrong — they solve different moments.

Mia: And both serve that same flow we were talking about: AI assisting the human in real time, rather than running off on its own. Which is exactly the philosophy of the next product too, and this one might be the most relatable thing in today's episode for anyone with an inbox.

Milo: ToneBird. It's a reply assistant for Gmail and Slack, on Mac and Windows. And here's how it works: it drafts your replies, and it remembers your relationships and your tone.

Mia: That memory part is what separates it from generic AI writing. A generic tool writes competent-but-generic emails. ToneBird's claim is that it knows how you talk to specific people — it remembers the relationship. So the draft to your close collaborator sounds different from the draft to a new client.

Milo: And crucially — you review before anything sends.

Mia: Nothing goes out without you reading it. Which is the human-in-the-loop answer to everything we discussed in the first half of the show. Naise claims agents running your social autonomously; ToneBird says here's a draft, you decide.

Milo: And honestly, for email and Slack, that's the right default. These are your words, going to people you have relationships with. Autonomy there is risky in a way that autonomy in video editing or content publishing isn't.

Mia: From communication, let's move to making things — and this next pair is notable mostly for what they don't have. No AI. No cloud, in one case.

Milo: First: Lightmeter, an iOS app. It combines two things photographers will recognize immediately: a light meter and a film camera.

Mia: A light meter — like the handheld devices film photographers have used for decades — to read the light and set your exposure. And a camera app that shoots with true RAW files.

Milo: True RAW is the differentiator. A lot of "film camera" apps are just filters over standard JPEGs. Real RAW gives you actual latitude in the file.

Mia: And the philosophy statement is explicit: no AI, no cloud. Everything happens on your device.

Milo: Pricing model: the camera is free, and there's a paid Plus tier. So you can try the camera and decide whether the meter justifies the upgrade.

Mia: For film shooters who carry their phone everywhere anyway, replacing a separate light meter with an app is genuinely practical — and the no-AI stance is a feature for that audience, not a gap.

Milo: Second one in this pair is for a different kind of maker: CodeSpotlight. It's a plugin for IntelliJ IDEA.

Mia: What it does is simple and visual: animated highlighting of selected code.

Milo: So you select a block of code and it gets an animated highlight, rather than the standard flat selection.

Mia: Who's it for? Tutorial creators and streamers. Anyone recording their screen while explaining code.

Milo: And that's a real pain point. If you've watched coding tutorials, you know how hard it is to follow someone's selection on screen. The standard blue highlight is easy to lose track of, especially in a compressed video. Animated highlighting makes the viewer's eye go exactly where the instructor wants.

Mia: Small tool, single purpose, clear audience. That's a nice contrast to the sprawling agent platforms earlier.

Milo: And that brings us to the last topic, which loops back to something we've implied all episode: all these products exist in an ecosystem, and makers need to make money within it. The last item is the Dub Program Marketplace.

Mia: The idea: one place to browse and apply to the top SaaS affiliate programs. And they've got recognizable names — Framer, Granola, Superhuman, and others.

Milo: So instead of hunting down each affiliate program individually — which usually means digging through a site footer or emailing someone — you go to one marketplace, see what's available, and apply.

Mia: For the audience listening: if you build an audience — newsletter, YouTube, whatever — this is the distribution side of the same ecosystem. The products we covered today, from Naise to Lightmeter, presumably all want affiliates. The marketplace is the connective tissue.

Milo: And it completes a nice arc: tools to make things, tools to communicate, tools to sell.

Mia: Let's pull it together. The through-line today is autonomy and its counterweight. We saw agents being handed real jobs — Pactto applying creative feedback, Naise running marketing channels, RankControl publishing content and chasing AI citations, Solid from an MIT, Berkeley, and Waterloo team that raised 6 million dollars to give agents their own computers, accounts, and budgets for long tasks.

Milo: Oh wait — I should mention Solid, since we skipped over it earlier.

Mia: Yes, let's do Solid properly, because it's the purest expression of the teammate idea. The pitch: agents that get their own computers, their own accounts, and their own budgets.

Milo: Own computer, own accounts, own budget — those three together mean the agent isn't borrowing your machine and your logins. It has its own workspace, its own credentials, and a spending limit.

Mia: Which matters for long tasks. If an agent is going to work for hours or days — research, iterating, buying resources — you need to scope what it can touch and how much it can spend. That's what "own computer, own budget" buys you: containment.

Milo: And the team: people from MIT, Berkeley, and Waterloo, with 6 million dollars raised. Early-stage, so again, claims and funding, not proven track record. But conceptually, Solid is the endpoint of the direction everything else is pointing.

Mia: And then the counterweight: AgentScore scoring those agents daily on five dimensions, Koreshield filtering what they see and do, Alexandria and GBrain giving them knowledge and memory, Opus 5.5 making the whole thing 40 percent cheaper to run.

Milo: Then the human-centered tier: Speechka translating your voice in real time, Linguo translating text offline, ToneBird drafting but never sending without you, Lightmeter refusing AI altogether, CodeSpotlight just making code easier to point at.

Mia: And finally the marketplace to monetize it all. If there's one open question hanging over the whole episode, it's the one we kept flagging: reliability. These maker claims — 40 hours saved, 80 percent cheaper, 21 percent better answers — are unverified. The measurement and security tools exist precisely because nobody trusts those numbers yet. That gap between claim and proof is where this category will be won or lost.

Milo: Exactly. Watch this space. Thanks for listening, everyone.

Mia: See you next time.