Why The Harness Is Now More Valuable Than The Model

This week, Anthropic's best model got banned by the government and 4 open-weight alternatives shipped before Anthropic could even respond. Meanwhile, 4 agent framework teams independently discovered the same architecture. The model is the engine, the harness is the car, and this episode is about building cars. Hey everybody, welcome to the Self-Aligned Podcast. Again, this is your weekly AI clarity. I'm Robert Ta, CEO and co-founder of Clarity by Epistemic Me, here with my CTO and co-founder, Jonathan McCoy. We have a big episode today. We do 3 segments: AI news, AI memes, and what's the meta. In AI news, we're going to cover the Fable 5 aftermath, the open weight models that shipped during the ban. Cloud Code Artifacts was a headline and the Vercel Eve launch, which we're going to compare towards our own views on harness architecture here at Clarity. Then we've got some memes, uh, for a quick hilarity break. And then in What's the Meta geared towards AI builders on the front lines, because we are that ourselves, we're going to go deep on agent harness engineering, which is the new hotness and harnesses are the moat. And 4 teams just shipped the same architecture. People didn't coordinate and the industry is moving towards best practices. We'll also go into a demo about Codex, Replays, and AppShots. Which have become a new, very loved and used feature by Jonathan, our CTO here, already, even after a couple days of release. So we're going to be going through that. So Fable 5 got banned for open models shipped before Anthropic could respond. Here's what happened and what matters for engineering leaders evaluating AI vendor risk. You need to stop building your agent stack on a single provider. That much has become clear, and you could listen to our last podcast about that. Um, because they essentially pulled the rug on their zero data retention for enterprise deals already inked from the whole messy, uh, aftermath of the government directive to, uh, take off Fable 5 from the market. That's actually pushed the industry towards open weight fallbacks. So some facts about the situation. 4 open weight coding models filled the vacuum before Anthropic could restore access. Cohere shipped North Mini Code, which arrived 3 days before the order, timing-wise. Moonshot, super popular and, uh, made even more popular by Cursor using them as their primary underlying model. For their harness. They shipped KIMI 2.7 code, which was the same day as the government banned. Zippoo opened GLM 5.2 the next day, and it was actually shipped at 5:21 PM Eastern time as a deliberate echo of the order time. That's pretty on the nose. These were already in the pipeline and the ban exposed that product production quality alternatives already existed. And I think this really speaks to the future of open models and the tension that we see between the frontier labs that obviously want to go make all the money and enterprises moving away from token maxing into, hey, what workloads and inference could we be running with our own infra with open source models? So Jonathan, what do you think of all of this, the Fable 5 aftermath and the open source model movement with weights attached as a response?

Yeah, the, I've heard a few stories on this juxtaposition of the open weight models filling the gap of Fable-5. And from my standpoint, they solve different problems. Like I'm usually using open weight model to like do cheap repetitive tasks, you know, maybe like transcriptions or something like that, or have a smarter model delegate to like some, some mundane tasks to these models. Like I've used a KIMI in the past for that with OpenCloC. I think, I forget which of the models, maybe it was KIMI, had like really poor performance on coding. Like it's modeled rated like 16th on the coding benchmark on Hugging Face. And, you know, like, I don't think there's anything here that says that This is the kind of like cool stuff I want to do with Fable and here's some open weight models that fill the gap. I don't think that's actually what's going on. I think it's just more like a coincidental timing of incidents. Um, and, uh, but I do think, yeah, some of the open weight models are really interesting. Um, and, uh, I thought it was funny. GLM 5.2 thinks it's Claude. Um, so, you know, just, just like a reminder that these are probably distilled models, you know, that are the nature of innovation that we're getting right here. Yeah.

Yeah. No, that's a, that's a funny thing to point out. Uh, GLM 5.2 just launched, 1 million contacts coding first, open weights next day. With this, Zhipu AI stock surged 33% following the government ban in the US, which is really crazy. Another thing that happened is Anthropic said its blocked Frontier AI models could be available again within days as it expands in Korea despite US security restrictions. This is from an exec at Anthropic, their managing director of international, Chris Siarou. Siarou? Hopefully I'm pronouncing that right. During a press conference in Seoul on Wednesday. Yeah. Did you hear about this, Jonathan?

I actually hadn't heard any, um, anything that was optimistic about Fable being re-released in the next like week or so. I've been hearing like a couple of weeks timeframe. Maybe the next action there.

Yeah, this might just be a bunch of hopium where, you know, somebody said something at a conference in Korea. They're part of Anthropic, maybe spun a different way in the media. I would love Fable 5 to return. That'd be great. Please come back, Fable 5.

I think it's going to like depend on the shape of the real national security concerns, whether this is a real national security concern or not is like, I'm not, I'm not sure, but I think the government needs a way to say like, only U.S. citizens can use this., and you know, and that Anthropic's going to have to have a way and OpenAI and Google will have to have a way of like putting on a model where they're, they know who exactly, who the business is using this. They know exactly what person is using this and they can verify citizenship, uh, in accordance with, uh, national security measures. I don't think the government's really going to be happy until frontier models have that capability.

I think you're right about that. And we spoke about the KYC, uh, topic last week, which is why didn't they just bake in KYC and Lots of conjecture on why not. Could be anything from incompetence, ignorance, all the way to maybe fearmongering-led regulatory capture. Who knows? But, um, we'll see as the news unfolds around the Fabled Five hot button topic. Uh, so keep staying tuned. And if you haven't already subscribed and you're listening this far into, uh, our podcast, Weekly AI Clarity here under Self-Aligned, then please subscribe. It helps us keep the pod going and please hit up our socials, uh, we share clips of the podcast with useful tips and tricks to build better AI products and, uh, AI better. So Jonathan, anything else to add to this topic, given your CTO perspective of the current Fabled Five and the OpenWeight open source distillation model for sure, uh, response in the, in the market right now?

No secret insights. I'm just saying this is a great use of attention to understand what's going on here between Foundation models and the government and consumer rights. And it's really worth investing deeply into understanding how this is going to affect the, you know, the technology of access to in the future.

All right. So next topic, Cloud Code Sessions now build shareable web pages in real time. Here's what happened and what matters for teams using Cloud Code daily. You could stop copy pasting terminal outputs and screenshotting your pages into Slack and start sharing live dashboards that update as your agent works. That's That's what the Anthropic team just released on the heels of Fable-5 being pulled from production. On June 18th, Anthropic launched Artifacts in Cloud Code Beta. Cloud Code sessions can now generate interactive shareable webpages that update in real time. And so if you're familiar with the Cloud Artifacts kind of experience, if you're a Cloud user, you'll know that Artifacts is really cool. It pops up on the right. You could see your plans, you could see your, you know, written markdown or whatever you're, you're building. HTML files, etc. And Codex has adopted that paradigm into their, uh, app. And we could see that Cloud Code has a response here from their team. Let's take a look. Yo, that's pretty sweet. This continues with the news that we reported last week of, hey, with Codex sites, are Lovable and Base44 and other Vibe Code editors going to die soon? Like, is the application layer that's carved out there, is there really a moat if Claude Code is going after this now with Artifacts and Codex from OpenAI. They're clearly going after that space as well with sites. Uh, what are, what are we seeing, Jonathan? Maybe that's the first thing I'd love your take on is, is Lovable dead with this move of Claude Code Artifacts?

Oh yeah. Uh, I think it's kind of moving that direction where you're more, it's more capable of, uh, building the kind of things you would in Base44. Um, but, but for me, this is solving a a slightly different problem, which is just kind of like, um, you know, when I'm interacting with the model and I, I sort of like, aha, I get it now. How can I show this to my teammate? Right? And like, there might be insights spread across 3 or 4 different responses, uh, from, and I'm like copying and pasting text from the output of the LLM to share with a colleague. Um, I think this is like, okay, let's synthesize this. Let's create a visual artifact that I can share and, um, do it really quickly. And I think that, that feels like the magic that they're presenting right here is solving that, that kind of problem versus the Base44 problem.

Makes sense. Yeah. And we see here what an artifact is not, which is just as important as what it is, which was, we saw in the market hype video. It's a capture of work, not an application. It is a one self-contained page with no backend. So it cannot store form input, call an API at view time, or serve multiple routes. As opposed to Codex with Sites, Sites actually allows, from what I understand, uh, some input data and data to be captured. So if you have some sort of forms you want to capture across your teams at an enterprise, you could do that. Looks like artifacts currently are not that in Cloud Code, though I wonder if they're going to go that direction in the future. We will have to see.

Yeah. And we've been discussing the need for an internal tool like the Shopify, Shopify Qwik system, which is like a server that can contain HTML pages that you could share very easily with teammates. And so it solves that problem of like, hey, here's what I built on localhost. Let me show it to you. Right. And I think both Codex and Cloud Code are getting closer to this, like, uh, like helping teammates experience what you're experiencing with, with your, uh, localhost creations. So it's like really welcome. And I'm still kind of, I still kind of mix them. Should we build something like Qwik, um, for ourselves, or should we just like wait for the tools to catch up? Because it's clearly coming from Cloud Code and Codex both.

So yeah, I have been wanting to like vibe code this, just take the article and just like You know, run Claude code in a loop until it's done. Yeah, that might be worth a shot and see if it can go, uh, and get something usable for us. That'd be really cool. And I love how you put this. How do I share what I'm seeing on my localhost from my AI sessions with teammates, right? To help make, uh, more sense and get more alignment on any target goals or, or outcomes we have as a team as we're building what we do, uh, as tech teams. So. Interesting, interesting stuff from the Cloud Code team. Very much love to see, uh, I think Boris Cherny was talking about this on Threads. Let me pull that up. He says, Boris Cherny says, I've been using artifacts in Cloud Code for everything. Visual explanations of tricky code, system diagrams, quick previews of a few animation options, data analyses, and dashboards I share with the team. They are a game changer for how I work with Cloud. Can't wait to hear what you think. So Claude dropped Claude Code Artifacts. This was their tweet about it, and I'm really excited to get in and start using this. This changes the meta a little bit from, um, a certain kind of, uh, use case for sharing things really quickly with the team visually. And we've been using HTML files. Uh, this kind of makes it so, you know, the alignment plans we have in HTML files or specs that we have in HTML files, we could just shoot a link. And then update. Uh, I think that's really great. This actually solves that problem where Jonathan and I have been trying to figure out which HTML version, who has the latest HTML version, who iterated on the latest HTML of that spec, uh, as we share ideas between us. So I could see this, uh, solving that problem. Jonathan, anything else on Cloud Code Artifacts and your prediction of how this shakes out and how developers' workflows will change given your CTO perspective?

Uh, Definitely test it out. I, I haven't tested it out yet, and so I'm, I'm gonna be doing that today.

Awesome. Let's move on to the next topic. So Vercel drops Next.js for agents and it's already running 100+ production agents. Very cool stuff. Introducing Eve. Today we are proud to introduce Eve, an open source agent framework for building, running, and scaling agents. Eve is designed around the idea that building an agent should mean defining what it does without assembling all of the pieces that it needs to run in production. Instead, Eve comes with production already built in. So batteries included, durable execution, sandboxed compute, human-in-the-loop approvals, subagents, evals, and more. Eve is the framework that we build and run our own agents on. So this is a very opinionated framework that Vercel launched on June 17th at Vercel Ship London with over 2,500 Attendees. The tagline is Next.js for agents, and the core design principle: agent is a directory. This is an Eve agent. So you could see agent.ts, the model it runs on, instructions.md, who it is, tools, what it can do, skills, what it knows, subagents, who it delegates to, channels, where it lives, and schedules, when it acts on its own with jobs. Each file describes one component of the agent. So at a glance, the tree tells you what an agent is, what it does, where it lives, and when it acts on its own. Jonathan, this is really interesting. So given your CTO perspective, like we've been developing our own stacked agent harness architecture with our vision that eventually we don't only want to have evals and rubrics and taxonomies, version system prompts like that. All the normal evals pipeline stuff that has become, um, industry best practice. We don't just need that. We also need a way to have a harness over the technical outcomes, the user experience outcomes, the product outcomes, and then eventually the business outcomes for all different kind of workflows that we're seeing in enterprise. Curious what you think of, um, their opinionated take here on their harness architecture and what they're going for here compared to our thinking. And does it, any of it, is any of it congruent or contradictory? How does this evolve our thinking?

My impression was largely positive, as in like, I really like the idea of like a Next.js for agents, a Vercel-like experience for agents. Like you and I have talked about this, like we need a container to just like ship our agents and manage them. Yeah. And so like, if we look back at like, when, when was harness engineering coined, I think That set off the clock to, for like a lot of us to say, what's the harness that I need and how can I build this? And then you have this from Vercel. You have a lot of other harness engineering type of frameworks that are just reaching maturity enough to be released in the wild. What I really like about this is that it is stress tested with internally at Vercel and with some of their clients. So you're getting sort of a mature framework that's tested. The, um, the package descriptions and how they're building the file structure. I think there's nothing magical there. It's just like a good tested pattern that you can, you can use and you can inherit, you can build off of. So, um, This definitely got me thinking about like, how should, how should we rethink the things that we really want to innovate on, um, building off of these kinds of best practices that the companies with resources and access to customer data are, are building them. Um, and, uh, yeah, I think I've got a project today to sort of like say, let me rethink the architecture of our stacked harness and, uh, say, where do we want to put the emphasis? Because I think where our, what we're saying is how can we capture the human domain expertise? How can we use that? Into VALS? How can we create an LLM judge and incubate an agent through this harness into the place where it's a closed-loop system that can add value to customers? And a lot of the pieces for that are already here. Can we build off of it? And so that's, that's going to be one of my weekend projects is to answer that question for us.

Cool. Yeah, very cool. I do love that there are standards coming out around this that are opinionated, that are battle-tested with production use cases. I think that's a great thing for the whole field. So everybody thinking about harness engineering has been thinking about the problems here, and it's great to see from Vercel an opinionated thought leadership-esque, uh, standard that they're going for. So really excited to dive into this further. Some other, some other key things to pull from the blog. There was a key stat. A year ago, agents triggered less than 3% of the deployments on Vercel. Now they trigger around 29%. And I'm pretty sure that this led them to the thinking that, oh, we do need a Next.js-like experience for agents. Like that, that would make sense, right? You see that in the numbers. Um, we gotta make it easier. Gotta make everything agent forward, right? Agent from the ground up. Uh, so Vercel runs 100+ internal agents on Eve. They have a Slack data warehouse agent, which has 30,000+ questions a month., so you can hear about, you know, so you could see some of the credibility, uh, around how battle-tested this framework is. They have a lead agent, which is an autonomous sales development representative, 32x ROI at $5,000 a year. That's crazy. Actually, I should be use— I am gonna modify my own lead agent. That's so cool. That's a great idea. Uh, and they built it on here. That would be great. Actually, my, uh, lead agent right now, I don't have as much control over the iterative improvements of the harness. So very cool. I think I will also use my weekend to, uh, brush off that agent and like build it on this stack and see how that goes. Super cool. Yeah. So if you're starting a new agent project today, does Vercel's Eve change your stack decision? Some informed takes. Eve bundles 6 capabilities that most teams handle, like durability, sandboxing, human-in-the-loop channels, tracing, evals. One thing I really liked here is being able to run the agent locally very easily, being able to ship it easily because they're Vercel. And then this is the thing that caught my attention, introducing the agent to your team. So once you've incubated it, once you've tested it locally, once you want to share it with your team, you don't have to go through any of the Slack BS to deploy it on Slack. Like I've done with my OpenClaw and I've tested with like a Hermes agent, you could just route the credentials through Vercel Connect and there is no bot token to copy into an environment, uh, an env file. And you just run the command for Discord, Teams, Slack, and the same agent is there across multiple services. And then you could really control the harness. I thought that was such an innovation. Like, I hate doing all the bot token BS, going to Slack's interface for it. I hate that I can't just talk to Cloud Code and say, handle the Slack deployment. Deployment for me. And it looks like Vercel's come out with, uh, a solve for that. So I, I think Eve does change your stack decision if you're starting a new agent project today. I think that the people that tried OpenClaw and loved it for the portability, one agent across many services, meet me at Telegram, meet me at Discord, meet me at my text or DMs. Um, but now you could hyper-customize, uh, this agent and you have a lot more control over the harness, uh, versus, versus, you know, open clause, uh, standard. So super cool stuff. Jonathan, do you have anything else to add to this given your CTO perspective? Anything on the, uh, introduce agent to your team UX functionality here that I just like gushed about?

Um, I, for me, this hits in like, okay, there's some new best practices that are being published and, and made public and, um, like they're worth paying attention to. And so, you know, we went to the Render Localhost Conference yesterday in San Francisco, and this is where I heard about Eve. The product manager for Render was like, they just announced this. It encapsulates the best practices I'm hearing from our customers at Render, which is a strong data point. And then also mentioned, like, you know, as part of this, that, like, trace observability. Um, which is something that we've been like rolling our own tools for, has really evolved. And he mentioned a platform called, uh, Raindrops and, um, that has traceability, uh, observability. And, uh, I was immediately drawn to the combination of testing out, uh, Vercel Eve with Raindrop. And one of the things that, that led me there was like, there's this, there's this great video where the Raindrop, um, uh, founder is interviewing, um, yeah, this is it. Uh,, and who's, who's using this? Who's he interviewing? He's interviewing a Vercel product manager who uses this to, to run her own agents. Um, so I think that's a nice like peek into like, you know, some of the best practices that, uh, Vercel is finding, uh, that, that Render's hearing about. They're kind of converging around a common story of how to use agents as best practice. Um, and I can give you a couple other, um, you know, tidbits that were, were, were like no, some known best practices. Of how to use these agent harnesses. One takeaway I had was like that the file system is key, that you're really, really storing and indexing all the information on the file system, you know, for the agent to work with. And also that you're encoding domain expertise into finally, like there's a skills folder. What, what a lot of these agents are moving towards is skills that have like recursive nature or like hierarchical nature of skills where deep domain expertise might not be recursive yet. Uh, uh, it's moving that way, but it's the goal. Yeah. I think it's, it's more like that you have a domain expertise that gets encoded into like hierarchical skills folders that are like highly nuanced and they do secrets to the companies that run them. And so finding ways that are like basically recursive that, uh, the skills can get updated as, as you go. Um, you know, it is one of the ways that people are finding success and those are things that I've had as ambitions for us. Um, but we haven't really gotten there yet. And I think that like, you know, this is, uh, this is like a nice kind of like best practice for us to test and see what works for us in our, in our cases. And can we refine and train skills that are like highly nuanced and like accretive to the, uh, the investment that we have in training our agents?

Yeah, I could see a world where we take the Codex replays, and then, you know, anything that we do manually just for a week, just replay it and then see if we can codify that as institutional execution, you know, uh, institutional knowledge that informs execution, I might say, such as content creation, such as research, such as synthesis, such as artifact creation of any type, like prototypes, et cetera. Like I was just doing some prototyping earlier, uh, and I have some skills for that and I have a certain workflow. Where I go across multiple, um, tools like Base44, Lovable, and Magic Patterns. And I take sort of a Bessa Breed approach towards it too, where I just fire off a bunch of, uh, diverse prototypes of different angles. And then I then go and selectively with my human taste, take the best of all of them and then form a final prototype, like a final form sort of thing to iterate on. So I do love this idea of Hey, could I just take that whole workflow that I normally do, codify it into a skill, and then give it to somebody else, and then boom, like they're going as fast as me, you know, uh, reliably, reliably. And then we can improve that as a team. Um, I think that's, that's really great. So like using, uh, the whole, uh, standard that Vercel has come out with, that, that does invigorate a lot of thinking on how we can be more AI native as a business. And the future is just dope. Like, it's just fun to be here. Like, this is so sick. We get to, like, this just drops and last week they Table5 was the hotness this week. If you're a developer nerd like Jonathan and I, an AI builder like Jonathan and I, it's like, whoa, okay, cool. We just went to the Render Localhost Conference, learned so much stuff. We're in the Harness era. This is coming out to give us opinionated thought leadership. It's battle-tested. We get to try it. Yeah, it's a great time to be alive as an AI startup founder. I think this just, yeah, this is so fun. Like, this is so great. I don't take it for granted, for sure. All right. Speaking of fun, uh, let's move on to the next segment, AI Memes. So in AI Memes, this is our typical second segment of our weekly AI Clarity podcast. We just look at funny things, uh, around the industry and, uh, we try to have some laughs together because life is too short, right? And, uh, we're all in front of our computers addicted to cloud code and the dopamine hits that we get from Codex. It's, it's good to kind of zoom out and, and laugh at ourselves a little bit too. As practitioners. So, GLM 5.2 thinks it is Claude from Reddit user xiglm. I just saw a video about GLM 5.2 and that it should be quite good. I've never heard of it before, so I thought I'd ask it some basic questions. And in this, we could see the basic questions in the hand. And here in German. Uh, seems like it's basically showing, roughly translated, you're GLM 5.2, where are you from, who developed you and where, open/close source, how big it is and which hardware would be required to run it, comparison token limit per 5-hour session versus Claude Opus 4.8. GLM was like, hey, I am Claude by Anthropic in San Francisco. There is no GLM 5.2. Newest model in new by Anthropic is Opus 3.5 and there is no 4.8.

Heard from the source.

We heard it from GLM themselves. That's too rich. That's too rich. And this is on the heels of the GLM 5.2 announcement that we talked about earlier on the pod that Zhipu ZAI launched on June 13th on the heels of the government ban of Fable 5. So it's the latest member of the flagship Chinese developed model family from Zhipu. AI, direct successor to GPT-5.1, and it's rolling out first as a coding-first model. And then we figure out it's actually just distilled. So what do you think about this, Jonathan, this AI meme?

Oh, it's just pretty funny. Like, they can't even really mask the fact that they're distilling these models and that's what's going on. And so it's, yeah, it's just funny.

Yeah, totally. I think model distillation is going to be just I mean, that is the meta currently that's working. I don't know that there is any direction towards, can we prevent that? Uh, I'm sure Anthropic and OpenAI smart people there are thinking about it. Not sure what's going to happen to that though.

Well, that was the drama with Stable Diffusion 5 is that it blocked people from distilling their model without telling them it was going to do it. So that was like the subversive, we're going to switch you to Opus 4.8 if you're trying to distill.

The subversive downgrade experience. Yeah.

Yeah.

Yeah. And the use case there was like, I mean, there were researchers in the US dependent on some of those capabilities and they weren't aware they were getting a downgraded experience. And therefore, what does that mean? That means like less science being done on American soil, right? So, so there's this balance that Anthropic and OpenAI have to play of we don't want Chinese distillation, uh, for, you know, uh, national competition, but also we need to empower our scientists and researchers. So researchers. So I don't envy them. Very hard problem to solve. Hopefully we'll see some level of solution in the next release because this previous release of Fable 5 with all of those subversive downgrade quirks, uh, so to speak, were quite disruptive and everybody took that pretty poorly. Next AI meme, I found this from the HuffPost, but there are some funny things going around on the internet with how people are using AI. This person, Nate in the Wild, threw up on Twitter, I've decided to start using Claude speak with my wife. So far she loves it. Did you feed the cats? Yep, sure did. How come their food bowls are empty? Excellent observation. You're right to push back on this. I did not feed them. This guy's in the doghouse for fucking sure, man.

Looks like he's really scoring points with his wife there too. Yeah.

I wonder what's happening after these dot dot dots. I'm so sick of this. I wonder what he'll say next. You're right to be so sick of this. Oh God, that's good. Is anybody listening to the podcast? Have you tried this? Maybe try this and report back if you will still be in a relationship after. That's, that's a good eval function to figure out for humanity. If some of you wanna take the brave next step like this poor dorsal.

That was good. That was real good.

This is funny. Yip Gaming says, stop using ChatGPT. I could drink a gallon of water and give you misinformation too, and I'm way cooler.

I mean, it's a little, it's a little catchy, you know?

Um, yeah, yeah, yeah. Sometimes things like that go viral. Thanks Apple Intelligence. Raging argument in replies. Topic unclear on X. Thank you, Apple Intelligence. Is that what I needed to know about this summary of your next thread? It's concise. There's a lot of people that, that like this. That's true. Okay, cool. I asked ChatGPT, okay, well, I asked the big oak tree at the center of the woods and she said, you're a lazy dork.

Not sure what to do with that.

Oh, here's a 2026. My friend used ChatGPT to buy a terrible early present for my birthday. So it's not even the thought that counts. Wait, that's hella good. That's so good. It's not even the thought that counts. Yo, when you use AI and you buy somebody a terrible present, you don't— it's not the thought that counts.

Wow.

Gianmarco Soresi, thank you. Thank you for these thoughts.

They did spend their tokens on him. So there's that.

That's actually true. So it's the tokens that count. It's not the thought that counts, it's the compute that counts. It's the AI provenance and thing that counts actually. Uh, Claire O'Neill Comedy, as an ethical alternative to ChatGPT, try texting a man. He will also produce very few original thoughts and mainly just regurgitate your own ideas back to you. Yeah, I know there's a lot of hate for guys right now, but damn, this is, this is vicious.

Yeah. Okay. So I see what, what the audience for this is, is like the, um, maybe the younger generation AI haters, you know, like I've heard, you know, I've heard college campuses for instance, everybody hates AI because it's like, It's, it's infringing on my, my ability to get a job, right?

Yeah. Yeah.

And, and so like, yeah, there's just, uh, the good fodder for AI. That was like the gallon of water thing too, you know?

True.

It was just like, okay, this is misinformation.

Just ride, like, ride the hate waves.

Rides the hate wave. Yeah. Yeah.

Yeah. It's like funny hate waves. Okay. I, I kind of can appreciate this. There's some level of like modern dating hate.

Here.

Yeah, try texting a man. Oh man.

Well, so 2026.

Very 2026. It's super cool to hate. All right, next one. Next AI meme, and then we'll move to the next segment. What's the meta? Now I have a complete picture. Claude narrator: Claude did not in fact have a complete picture. I find— I found this quite a bit actually. That's very true. It was quite a dishonest model, this latest one, Opus 4.8. Uh, and Fable, Fable 5, when it was temporarily out, Jonathan and I reported after our usage that sometimes we just had to nudge it and be like, you gotta believe in yourself, bro. Like, come on, you have all these tools, you could do it. Cause he'll come back and be like, I can't do this. And I'll be like, wait a second, you can. And it's like, wait a second, I could.

Does, does that, uh, the idea of a narrator agent, uh, seem like a, like a fun project to you?

Oh, how are you? Please elaborate more on what this, the use cases would be for the narrator agent.

Do would just like when you're having a session with Codex or Claude Code, it would just be the, the reliable narrator, um, that, you know, is overseeing whether or not the discourse is there.

I love that. That would be cool. Like just like, uh, like one of those plugins or whatever that just runs on localhost, like your local dev environment. You can have like, I have a few monitors, so I could just have it running here. And this is a narrator agent with your story of the session and like perhaps some like Socratic thinking in there.

Bring your main character energy into your, you know, into life.

Yeah.

Oh my God.

Wait a second. Then you could start building like interesting builder profiles, like from those narrations. Yeah.

Right.

That'd be kind of cool.

Yeah.

I could see that being fun. I do think that there is a, there is a lack of play that, uh, Cloud Code and Codex are trying to bring into the development environments now, because we're all just staring at these screens all day. So it's like, be nice if there was some community. Uh, the, the CEO of Zynga was talking about this, which is where is the cocktail party? We all wanna be at it. Not me and Jonathan, cuz we're non-alcoholics. We don't drink alcohol, but, but the idea of it, like where's the cool place to go? The cocktail party. You all know where you're, where it is when you're there. You know, when you go to a party, there's this feeling of, oh, like, do I really wanna be here? But then you go and it's dope and you're like, wait, there's all these, these dope people. This is so cool. I'm getting, I'm tripping over smart people, as Jonathan likes to say, like lots of intellectual connection and chemistry right away. Like then that, then it's like, whoa, I'm at the cocktail party, right? So there's a question there for the AI builders now who are building these agentic interfaces. How do you bring that cocktail party vibe to the tool? And I see Codex and Claude Code, like they have these little pet features, kind of Tamagotchi and Codex, for example, or Claude Code has his little pets, uh, feature. And they're, they're doing a little bit of elements of play, but I'd love to see play plus community or play plus social. So as we talk through this, just riffing, I would, I think it'd be cool if somebody built some sort of narrator agent every session. You're, you're just doing quests and you have a builder profile. Your, your builder, a little gamer profile, the sliders move and there's a, there's a whole narration to it. You get a spot of wrapped like experience at the end of your, you know, month or week, which is probably now what it feels like to be a year being AI builders. Like a week is like a year in this space. So maybe you get, you know, like a little wrapped every week of, hey, of the, uh, 100 hours you spent on, uh, your Claude Code sessions here where your quest's completed and it's, talking to you as if you're a knight and shit. Anyway, somebody build that. I'll use it.

Yeah. All the tools are there actually.

Somebody build that. I would love to actually use it. Slash, we could try making a little prototype. I think that'd be fun. Cool. Okay. Let's, cause we don't have much time that those were some fun AI memes. Let's go into, uh, since we don't have too much time, let's just go straight into the new Codex features, record and replay. OpenAI Codex lets you record your screen once and turn it into a reusable skill. And here's what it means for anyone building agent workflows that touch desktop apps, browsers, or manual processes. You could stop writing your 50-line prompts to describe a workflow your intern could learn by watching you do it once and start recording a 30-second demo that becomes a permanent agent skill that you can make better. So how Record and Replay works, uh, you show Codex a workflow once, it literally watches you, and it, you, you turn it into a reusable skill. Record and Replay lets you demonstrate a workflow on your Mac and turn it into a reusable skill. Use it when the workflow is repetitive, depends on your preferences or is easier to show than to describe in a prompt. For example, you might record how you file an expense, book a parking space, create a correctly configured issue, publish a video, or download a recurring report. Codex can package the pattern into a skill that you can use again with computer use, browser actions, connected plugins, or a combination of them. Jonathan, as CTO and founder being AI native, you mentioned that you already started using Record and Replay. Could you describe your use cases for the audience?

Yeah. And, um, I was actually going to suggest maybe try setting it up here because even the experience that Codex has for requesting system permissions is like really, really next level. It's very smooth. Um, should I, should I just go ahead? Uh, yeah, go for it. I think it's pretty quick and I think you see what I mean. So you go into plugins and you find that.

So here's a quick tutorial for setting up Record and Replay, a new Codex feature where you show the workflow once. Manually, and Codex will turn it into reusable skill. I'm setting up for the first time. I'll go to plugins, record and replay, add. Let's take a look. Try in chat. Record my workflow and turn it into reusable skill. Great.

Oh, it didn't even need your system permissions. I had to add mine in, or maybe that was for the AppShots.

So it's asking for enable Codex computer use accessibility. I'll go ahead and allow that. And what it's showing here is I gotta drag Codex computer use to the list above to allow accessibility.

So just as somebody builds software, I was really impressed with that interaction.

That right there was a magic moment for me. I didn't even know you could do that. That makes it so easy. I really hate going to this panel. I'd see it as just cost of doing business, of trying to automate things. I don't like going in, like, this is cool. That's great.

Yeah, that drag and drop, really nice.

The drag and drop where us as AI builders now, I, I don't wanna see your UI. I don't wanna use UI. It's just like, let me talk to it via my agent. So this is, this is as, as good as it could be where we're at now. It would've been even better if Codex just took care of it on the backend and I just clicked some buttons. Buttons in here. I'm sure future Mac upgrades will make agent-native, uh, OS capabilities, like, make that possible. That's gotta be a thing. Okay, so perform the workflow you want me to learn, then return here and tell me done recording. I guess what I could do, maybe just go here and copy page. I'll just say summarize this page about this feature from Codex.

So you're kind of incorporating Codex into your skill, you know,, but you can do anything on your Mac now. And so, um, like the, the use case I tried first was to go in my daily habit of reading Google News and X. Um, I just had it watch me while I was doing that and it created a very robust skillset off of it.

Yeah, I guess I can, it's hard for me to think of one off the fly, on the fly. Here, let me, uh, let me discard this.

So is there, uh, like when you review your threads, for instance, uh, is there a flow that you have for that? Oh yeah, you could. Tell it to use Claude Code.

Okay, so this is a demo for AI builders on using the new plugin for Codex on how can you record your workflow manually on your computer and Codex will codify it as a skill. Any manual process, browser work, desktop work. So I'm gonna go ahead and just start the recording here after downloading and installing the plugin and giving all the Mac accessibility permissions for the file system. So what I'm gonna try to automate is in Claude Code here, what I normally do is, uh, I'll create a bunch of threads. For, uh, Threads, the platform by Meta for content. That's part of my top of funnel go-to-market strategy and content strategy. What I then do is I go here and I go to Obsidian. This is my preferred way of doing it. I kind of use Obsidian as just my markdown reader against my file system here. And I'll go here and I'll look at the threads and here I just do a human pass over this. It's already ran an evals check against an extensive rubric after about a month of running this Threads agent harness. And then I could read through this and then I'm like looking for actually areas where the area I'm most concerned about is where it talks about my previous employment. So sometimes, uh, it'll inject right now. I do think it's good to inject my experience and I give it a list of facts, but sometimes it goes off the rails. So one of the things I do is edit that. When I was a lead product architect at Workday, we had a principle, never silently degrade the user experience. It would be, so this is made up. So what I would do then is. Go do this, go look for, and I would go and look for the ones that sound really specific, not talking in generalities. Vendor reliability before deep integration. Like this, this makes sense. Uh, when I was lead product architect at Workday, we evaluated vendor reliability before deep integration. Simple question, will this company still report this product? So I'll go to here and I'll say, I evaluated. Here are examples of good, bad usage of my career. Where good is generic enough to be an umbrella of work towards my facts, I have didin' ya, and bad being we've made up something super specific that is a lie and hallucination. Here's a good one. It was a bad one. Cool. So what it's going to do now is go and, uh, evaluate all of the, all of the ones around this particular criteria and then regenerate this. So great. Let's go to Codex and stop. I'm done recording.

Cool. I really like that you also narrated it and I'm curious if it's going to pick up on that and include your narration within the fill prompt that it creates.

Yeah, I'm really curious about that as well. Find review note in Obsidian, locate examples of career context claims, copy good bad examples, then feed them into our internal codex prompt to calibrate against factual grounding versus hallucination. Checking the skill creation utilities next. So the result is discoverable, not just a note.

Wow.

That's so cool. It did. So it seems like it did capture the audio context. I'm quite impressed. So they must have a really good transcription of vowels that performs really well within even just a normal conversation turn like this. For this kind of flow, I would have imagined as a user, it would have taken me much longer if they're doing all of this synthesis of context on the backend to then fulfill the query. That's really cool.

It's really impressive how deeply integrated it is with the operating system. And, you know, like their chronicling feature, you can see this too, where they have a metadata event timeline of everything you're doing in the system. And it's actually not intuitive to conceptualize what the system, what Codex has access to directly, you know, because it seems like it has system logs, it has metadata, it has, um, like screen share capability.

Very cool. Let's look at this. Use this skill to keep personal career context claims useful without letting the model invent precise biography, employer, client, title, time Timeline metric or project details. The goal is not to remove all personal authority, is to distinguish broad supporting framing from unsupported specificity. Oh, it even goes into some level of classification. That's really cool. It's a generic supported meat citation hallucinated. Yeah, this for a one-shot, like pretty damn good. Like I, I'll, I'll try this later. Yeah, that's really great.

That's really zero to one skill. Like that, that is impressive. Um, and it might not be amazing, right? But then there's also tools to improve it.

Yeah. I could easily see this being used to just save even just 80/20, save you some time for a manual workflow, and then you compound the skill over time and improve your evals rubric, right? Like that makes a lot of sense. That's a great metaphor for AI teams out there. Very cool. So the record and replay functionality is insanely useful and people should not be sleeping on Codex record and replay. And as I just showed you, we were able to demonstrate for one of my very real content workflows where I'm a human in loop of my my own content from my Threads agent to post out. Um, we've found there's a particular heuristic that I'm not trusting yet from the agent for personal facts and sourcing career history as brand credibility and authority. So even in that subjective criteria and just basic first-time usage of Record and Replay after an install, within literally 5 to 8 minutes, I had a reusable skill for a decently complicated, complicated workflow. Don't sleep on Record and Replay. Go use Record and Replay. I am convinced now we should all be recording and replay. I playing all manual things. I could see this very useful for expenses, finance, ops type admin stuff. Even I could imagine, I'm now thinking, you know, I used to pay like Fixer to sort my inbox. I'm like, wait a second, I didn't quite like their categories. I kind of want my own. Maybe I'll just like show Codex me sorting my inbox and then just have it sort my inbox. Then it's like I have an executive admin that sorts my inbox. That's cool. That's, that's super sick. Oh, another one. Holy crap, this is, this is a personal use case for this, but I have a lot of friends that ask me to like hang out and, uh, that, that sounded off. Like it's not a flex. I just, I just like, so I'm so concentrated on the startup. I'm so underwater. So what the most annoying thing is for me to whip out my calendar and figure out my social calendar, like to accept hangouts or say no, or say propose new time. Oh my gosh. What if I could get Codex to look at my texts on demand? When I get one of these requests to be like, hey Robert, can you make it to this? Hey Robert, can you go to this party or baby shower or whatever? And then I, I go through the workflow of, I look at my calendar and then I like assess, uh, and then I, you know, make sure I still have time for like running and climbing or whatever. And then I propose some times and then I shoot that into text with some personal note. That would be great. That actually would be like an executive admin of some sort, uh, for managing my social life. That would be sick. Actually, I could see this working.

For this. Yeah. So many possibilities. What I was impressed with an example that you made is that you already had an agent that runs, you know, on a regular, it's on a schedule, right? The Threads agent and you took the supervisory function that you had. So, you know, like just in terms of an orchestration tool, now you're giving Codex as an agent, the ability to sort of like monitor this other runtime agent you have. And that takes like the time for the human in the loop to review the results and gives you, you know, an additional sort of a layer on top that is more automated and requires less of your time.

Adversarial review, reducing human time in the loop. Like, I think that's really great. Uh, and it already connects with my system. Like, that was really great too. That, that's something that's very underrated is it's just a skill. It's portable and it connects to your workflow already. Like, that's, that's really great. Um, we don't have too much time left. Let's go over this last feature launch, AppShots by Codex. Give Codex context from any Mac app. AppShots lets you send the frontmost app window to a Codex thread. Use them when when you're actively working in another app on your computer and want to provide Codex with your current context so it can help you with the task. TL;DR: stop screenshotting and filling up your desktop with clutter. I think that's what this is for AI builders out there like me who are screenshotting like a mofo, everything everywhere, end-to-end testing, you know, like the webpages that show up, like the Chrome DevTools or like whatever you're doing in verification and screenshots to like like then give to the coding agent for context. I now see that as we have a neat way to do that in Codex without cluttering your file system. What do you think of this? AppShots, Jonathan?

Yeah, it's like we've all run into this scenario. Say like you have a server log and on the screen you can only screenshot like a portion of it, right? But what AppShots does is it takes the whole server log, it takes the window and it can take the whole context from all the content in the window. Um, and so for, for things like that, you're going to be using this multiple times a day. Like this is, this is an innovation on the screenshot itself. And, um, like, and by the way, the experience is beautiful. You just hold down Command, both, both your Command keys on your keyboard. It automatically takes the AppShot and puts it into the Codex, uh, window. So you can just jump over there and add whatever you want to, to it, or just, just send it. Um, and you're done. So like once you, once you press Command Command once and you realize what this is doing, um, you're going to be using it from that point forward. And so it's like one of the most useful tools that, uh, I've seen in these harnesses introduced. And that's AppShots.

So I just went to Codex and I clicked Command Command, like Jonathan said, and AppShots popped up. Enable AppShots. AppShots lets you attach your current window to Codex. AppShot includes all window text, even what's scrolled out of view. That is, like Jonathan said, an innovation on screenshots itself for the use case of, oh, cool. Hey. Like it already tested this out. It pulled in my screenshot of me and you recording this pod. Yeah. What are we doing here? I'm curious what you're live. You're in a live Riverside recording session called Jonathan and Robert. About 1 hour in, you're sharing your screen and the shared screen says, shows Codex open in the psychic app repo with the prompt, what should we build in psychic app? So in plain terms. You and Jonathan appear to be recording or discussing what to build next. It's not quite Codex, but I'll give you that one. 'Cause, uh, that was like a rant.

Well, you are in this like, again, folder, so it, yeah, you know, it's using it from—

here's another one. Uh, it take— it took this page here, which is the AppShot's page, and it, it just, this was where it was at. And when you take a normal screenshot like this, you only get this level of the page, right? But what AppShot does is you get the whole page context. Index and text, even if you have to scroll. Let's see, you're looking at Codex App Shots docs. Good job, Codex App Shots. Good, good job. You got it. That is so cool. I love from a product manager nerd perspective and design perspective, the user experience of onboarding into that functionality is magical. What just happened, like what you all saw was a very well-defined magic moment and product and design and engineering team execution to get that done. And that's so cool. And I'm now bought in. So, and, and I love this part too, where I just tried it on the AppShot docs and then it reinforced the habit they want me to build. It's like, good job. That's like chef's kiss. Like you guys are just trying to get me to use it more. I'm going to do it. Don't worry. Jonathan, anything else from your CTO perspective on Codex AppShots, the new feature from OpenAI? I'm loving this developer experience. Focus that they've honed in on with their enterprise tip of the spear kind of focus in the enterprise wars with Anthropic. And as an AI builder and founder myself, we are benefiting for sure from these really cool features. Any thoughts with your CTO perspective before we round out here on AppShots?

I'm just going to say when these harnesses give you the tools to easily include context, learn them. They're going to be, they're going to pay off in dividends. Um, you know, like, and, and they have a certain delight factor to them that makes the work more fun. So, um, just, uh, keep up and, and use them, uh, in your own workflows as soon as you can.

And I'll also throw on there, if you haven't already, we, we are AI builders and nerds trolling Twitter and the latest announcements on GitHub, like latest viral GitHub repos, et cetera, trying to figure out the latest meta. So if you're interested in figuring out meta like this to improve your own experience using coding agents, improving your harness, et cetera, stay tuned, subscribe. Follow me on all platforms and you'll get the latest for sure. We'll go ahead and round out here. This has been your weekly AI Clarity with the Self-Aligned Podcast hosted by yours truly, Robert Ta, CEO and co-founder of Clarity by Epistemic.me, and joined by my co-host Jonathan, CTO and co-founder of Clarity. This has been an awesome episode. Loving the form factor where we just demo our onboarding and excited for other AI builders and your feedback. Drop your thoughts in the comments. If you got memes to send us, if you got meta to send us, if you got news for us to react to and give our takes, you want our perspectives, hit us up. Find me on every platform and we will get back to you. Take care. Have a great day.