
1119 | Agents Everywhere: This Week in AI Tools
Show notes
Frontier model releases, agents that code and test, agent-building platforms, voice and communication AI, plus marketing and creative tools — a fast tour of what shipped this week and why it matters.
Timeline
- 00:00:04 Opening
- 00:00:40 Frontier and efficient model launches
- 00:02:41 Specialized models: security and real-time decisions
- 00:04:12 Agents that code, test, and fix themselves
- 00:06:20 Seeing what agents break: UI, QA, and review
- 00:08:29 Building blocks for agent workflows
- 00:11:02 Team and desktop agents that persist
- 00:12:34 Agent infrastructure and design systems
- 00:14:05 Voice and communication agents
- 00:15:07 Marketing, creators, and monetization
- 00:16:45 Creator and design utilities
- 00:17:44 Closing
Related links
- GPT-6.2 Astra and Sol
- Qwen4-Max
- Tencent Hy4 flash
- MAI Cyber 2 Flash
- uRun
- SoreQen Code
- Polylane for CI
- Replay QA Self-healing Mode
- UI Verify
- Review Box
- KaDeep AI
- Toone
- Mermaid Flow
- dusk bot
- Dippr
- S.Space
- Incredible Agents
- Dream by Smart Jelly
- AI Canvas
- canfit
- Waldo by Wherobots
- Pitch Avatar HR Suite
- Bluedot Phone Calls
- Arcade
- ICONS
- Introzy
- Diagramly 2.0
- lepoly
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.
Transcript
Mia: Welcome back to the show, everyone. I'm Mia.
Milo: And I'm Milo. Today we've got a whole stack of launches to get through, and the thread that ties almost all of them together is agents — agents that write code, agents that run overnight, agents that need data and design systems and memory of their own. Plus the models underneath them are splitting into two camps, which is kind of the story of the day.
Mia: Right. So let's start where the day starts, with the models themselves, because the launches really do split into a frontier option and an efficiency option almost every time now.
Milo: OpenAI is the clearest example. They released GPT-6.2 in two flavors: Astra, which is the frontier model, and Sol, the efficient one. Sol's API pricing is two dollars per million tokens in and ten dollars out. So if you're running an agent that makes thousands of calls a day, that efficient tier is the one that decides whether your product math works at all.
Mia: And the same split shows up on the open-weight side, sort of. Qwen4-Max is built on a new mixture-of-experts architecture with 2.4 trillion total parameters, but it activates 71 billion per token — so it's a big model that only wakes up the parts it needs. It has native vision and a million-token context window. The catch, and this matters, is that the weights are not open despite the Qwen lineage people associate with open releases.
Milo: Then Tencent's Hy4 flash makes a similar trade in a different spot: 224 billion total parameters with only 12 billion active per token, a million-token context, and input at seven cents per million tokens. They're saying they plan to open-source it, but the weights haven't been released yet. So it's a promise, not a fact, and we should hold it that way.
Mia: I think that's the honest summary of this whole cluster. The frontier-versus-efficient split is now basically the default product strategy. Every lab seems to need both a headline model and a cheap workhorse. And the open-weight question is genuinely open — Qwen4-Max is closed, and Hy4 flash is only "planned" open. Whether the cheap models get cheaper to host yourself, or only cheaper via API, is still undecided.
Milo: Okay, next slice: models built for one very specific job. Two launches here that couldn't be more different in purpose.
Mia: Microsoft has MAI Cyber 2 Flash, a security model with a million-token context and a reported CyberGym-E2E score of 92.3 percent. But — and this is the big limitation — it's restricted to use inside Microsoft's own security products. So it's not something you can take and wire into your own SOC. The benchmark number is Microsoft's claim about the model; whether it holds up in your environment isn't something we can verify from here.
Milo: And at the other end of the speed spectrum, uRun's Sense is a real-time decision model for voice and video agents. Its whole job is answering the turn-taking question — when does the agent speak, when does it wait — and it does that up to 12.5 times per second. That's a very narrow, very timing-sensitive task that general-purpose chat models handle badly.
Mia: What links these two is specialization. We've spent a couple of years assuming one big model would do everything, and now you have Cyber 2 Flash for defense and Sense for split-second conversational timing. The efficient generalists we just talked about sit in the middle. The question worth watching is whether these specialist models get absorbed back into the big ones, or whether a specialist layer becomes permanent infrastructure.
Milo: Let's move to agents that actually build software now, because there were several launches and they fit together as one loop.
Mia: SoreQen Code is the most self-contained one. It's an autonomous coding agent that checks its own work — it edits code, runs tests, and verifies things in a browser. It ships as a desktop app plus an open-source CLI, and it works with 18 different models. All of that is the maker's description, so treat it as claims, but the shape is interesting: write, test, look at the result, iterate, without a human in the loop for each step.
Milo: Polylane for CI starts further downstream. It reads failed CI logs on your main branch or merge queue, finds the commit that broke things, opens a fix pull request, and then keeps iterating on its own PRs. It works with GitHub Actions. So it's not trying to write features — it's trying to shrink the time between "main is red" and "main is green."
Mia: And Replay QA closes a different gap entirely — the gap between what tests catch and what real users hit. It captures sessions from frustrated users, automatically reproduces them, does root-cause analysis, and hands the result to coding agents to fix. So a user rage-clicking through a broken checkout can end up as a fix ticket without a human triaging screenshots.
Milo: Put those three together and you get a loop: CI breaks, Polylane finds and fixes the commit; a real user hits something the tests missed, Replay reproduces and diagnoses it; SoreQen does the edit-test-verify cycle inside that. The theme is agents closing the loop from bug to fix with less human babysitting. The open question for all of them is trust — how much of that loop you're willing to let run unattended, which is exactly what the next group of products is about.
Mia: Right, because once agents write code, someone has to see what they broke. Three tools here that split neatly into pre-merge and post-merge.
Milo: UI Verify gives coding agents visibility into every interface their changes touch, and automatically fixes visual regressions it finds. For open source it's free with 50,000 snapshots a month. The pitch — and again, this is the maker's pitch — is that an agent changing backend code often can't see the UI it just affected. This is trying to be its eyes.
Mia: On the pre-merge review side, Review Box is a desktop app for GitHub pull request review. It has reviewed-tracking, so you can see what you've already gone through, plus code intelligence and integration with AI assistants. That's more of a human workflow tool — it's making the review process itself better rather than automating it away.
Milo: KaDeep AI sits across both sides. It's an agent that unifies QA, localization across more than a hundred file formats, and accessibility audits, and it produces an evidence-backed go/no-go release gate, delivered through an open-source CLI. So before anything ships, you get one verdict with the evidence attached.
Mia: I like the way these three map onto the problem. Review Box is the human eyeballing the PR before merge; UI Verify and KaDeep are the automated checks that catch what eyeballs miss. If the agents from our last segment are going to write more of the code, this category is the safety net. What none of them solve is the final judgment call — KaDeep gives you a gate, but a gate is only as good as what it checks.
Milo: From checking agent work to building agent workflows — several infrastructure-ish launches here.
Mia: Two of them are visual or natural-language workflow builders. Toone lets you assemble agent workflows from natural language, where each node is either a model chat thread or a plain non-AI step, and you can pause a workflow and resume it later. That pause-and-resume detail is more significant than it sounds, because real workflows run for hours or days and you don't want everything lost on a restart.
Milo: Mermaid Flow takes the graph approach: you build an executable diagram of your workflow, connect MCP tools with one click, and every run leaves a step-by-step trace. That trace is the feature I'd underline — when an agent pipeline misbehaves, being able to see exactly which step did what is the difference between debugging and guessing.
Mia: Then two launches about what agents need beyond a flowchart. dusk bot is a free, open-source multi-agent desktop app where each bot gets its own computer and its own memory, and you can connect your existing ChatGPT, Claude, or Grok accounts instead of paying for new ones. And Dippr is a data platform for agents — it does federated queries across sources, scopes authorization per agent, and hosts the apps agents build.
Milo: That data piece is the unglamorous but probably load-bearing one. Agents are only as good as the data they can reach, and "which agent can read which data" is a real security question that most demos skip. Dippr is trying to make that a first-class concern rather than an afterthought.
Mia: Together this group is the plumbing layer. Workflows, observability, compute per bot, data access. None of it is flashy, but if agent products are going to graduate from demos, this is the stuff that has to exist. And it connects directly to our next topic, which is agents that stick around long enough for that plumbing to matter.
Milo: Right — persistent agents. Three launches with a shared idea: an agent that isn't a one-shot command but a collaborator that keeps existing between tasks.
Mia: S.Space comes from the AutoGen team, and it's a shared workspace for teams and their AI agents across different vendors. The two features worth naming are shared memory — so agents and humans see the same context — and approvals, meaning an agent's actions can go through a human sign-off. That's an honest acknowledgment that full autonomy isn't the product yet.
Milo: Incredible Agents lets you create agents by voice, and those agents run tasks directly on your Mac or Windows machine around the clock, proactively finding work that can be handed off to them. It's free to try. And Dream by Smart Jelly is a Mac agent aimed specifically at overnight-friendly tasks — it works through them while you sleep and reports the results in the morning, and it shares memory across tools.
Mia: The direction in all three is the same: agents moving from single-task helpers to persistent collaborators with cross-tool memory. And they raise the same unresolved questions — an agent running for hours on your machine with shared memory across tools is powerful and also exactly the thing you'd want approvals and scoping for, which is why S.Space's human sign-off model is a useful contrast.
Milo: Now the layer underneath all of that: infrastructure and developer tooling for agents themselves.
Mia: AI Canvas ships a complete React design system plus an MCP server for AI agents, including 820 icons. The detail that stands out is something they call a done-gate — rules that get enforced so an agent can't declare work finished prematurely. That's a clever idea: the design system isn't just assets, it's guardrails the agent literally cannot bypass.
Milo: canfit solves a much smaller but very real problem. It's an open-source Rust CLI and TUI that tells you whether a local LLM or an app will fit in your machine's current memory, and suggests which processes you could safely close to make room. If you've ever tried to load a local model and watched it fail at 90 percent, you know why this exists.
Mia: And Waldo, from Wherobots — the creators of Apache Sedona — is a geospatial AI agent in public preview. You ask questions in plain English and it turns them into maps and data pipelines. Geospatial work has traditionally required real expertise in specialized tooling, so a natural-language front door to it is a meaningful change, with the usual caveat that preview-stage claims deserve preview-stage skepticism.
Milo: Okay, voice and communication — two launches where AI takes over routine spoken workflows.
Mia: Pitch Avatar's HR Suite uses conversational AI avatars to automate onboarding, training, and HR communication, in more than 70 languages. The use case is repetitive — the same onboarding explained for the fiftieth time — which is exactly where scripted avatars make sense. The open question, as always with HR-facing AI, is how employees actually receive it, and that's not something we have evidence on.
Milo: Bluedot added phone call records: transcription, summaries, and to-dos, integrated with a dialer. Two honest limits: it's outbound calls only for now, and it's limited to certain countries. The pattern across both of these is AI absorbing the routine layer of spoken communication — the notes, the summaries, the follow-ups — while the judgment stays human, at least for now.
Mia: Now the marketing and monetization cluster — three products aimed at people trying to earn money from content and brand work.
Milo: Arcade is an AI video generator built for marketing teams. It creates editable, branded videos from prompts or from sources like Notion, and you can publish directly from it. The "editable" part is the differentiator worth noting — a lot of AI video tools hand you a finished clip you can't change, and marketing teams live in revision cycles.
Mia: ICONS is a marketplace with two modes: brands can book real creators for sponsored posts, or license a creator's AI likeness. Creators set their own rates and approve the scripts and the final cuts. That approval structure is the interesting part — it puts the creator in the loop on anything their face or voice is attached to, rather than licensing being a one-time signature.
Milo: And Introzy handles the money side of affiliate marketing. It tracks referrals from first touch all the way to commission payout, it's free to start, and it has an AI assistant named Kai that flags stalled referrals. The problem it's aimed at is real — affiliate pipelines leak in places nobody watches, and a referral that stalls silently is money nobody notices losing.
Mia: Together those three cover the arc: create the content with Arcade, monetize your likeness with ICONS, and track that the money actually arrives with Introzy. For all three, the claims are the makers' own — we haven't seen the outputs or the payouts.
Milo: Last stop: two small, focused creator tools, and honestly a nice note to end on.
Mia: Diagramly 2.0 turns typed text into hand-drawn style diagrams, and this version adds review links, comments on individual shapes, and slide presentations with speaker notes. It's free with no sign-up required. The hand-drawn aesthetic is a genuine choice — not every diagram should look like a polished corporate slide, and for explaining ideas casually, that style is often friendlier.
Milo: And lepoly is a browser-based digital asset management platform for 3D assets, built by the Sketchfab founders, self-funded. That last bit matters — 3D asset management is a hard, unglamorous problem, and a small self-funded team choosing to attack it suggests they think there's still room for focused independent products in a market full of platforms.
Mia: Which is a good closing thought for the whole episode, actually. Most of what we covered today was agents — agents coding, agents testing, agents running overnight — and almost all of it is early, with claims we can't yet verify. But underneath, the pattern is consistent: the models are splitting into frontier and efficient, the agent tooling is moving from demos to infrastructure, and the safety layers — approvals, gates, traces, scoping — are being built at the same time, not after.
Milo: Right. The verdicts are still out on basically everything we discussed, and we'll see which of these hold up as people actually use them. Thanks for listening, everyone — we'll catch you next time.
Mia: See you then.