Transcription
Today we're discussing the dawn of the agent era. Welcome back to the AI daily brief.
As part of my ongoing collaboration with KPMG, each month at the beginning of the month, I'm doing a bit of a retrospective show to discuss and try to summarize the key themes from the month that came before. Now, obviously, when it comes to January, we're actually a couple days behind already firmly into this month. And of course, the reason for that is everything that went on with Moltbook and how that plus the SpaceX XAI acquisition nudged stories up. But in many ways, those stories which nudged the January recap back are perfectly aligned with the themes that shaped the first month of 2026.
I'm calling this episode the dawn of the agent era. And right from the beginning, it was clear that something had shifted. Now, what was interesting about this shift is that there was a bit of a lag between when the capabilities came online and when people really recognized it. Midjourney's David Holtz had the representative tweet when he said on January 3rd, "I've done more personal coding projects over Christmas break than I have in the last 10 years. It's crazy. I can sense the limitations, but I know nothing is going to be the same anymore."
So many people had the same experience of going back home, slowing down a little bit, and having a chance to actually see what you could do with models like Opus 45 and Claude Code and GPT52 codecs. It turns out the answer was a lot and folks came back convinced more than ever that we had hit a fundamentally different period in the history of AI coding and by extension the history of AI. Vibe coding, which as a term just turned 1 year old the other day, over the course of this month shifted from the thing that you use for prototyping to just the thing that you do.
This shift itself was never better embodied than when Anthropic released Claude Co-work, the Claude code for everyone else, and shared that it had been built in 10 days, pretty much all by Claude Code itself. Throughout the month, we got articles like this one from Maxwell Zev, "How Claude Code is Reshaping Software and Anthropic." We got commentators outside the tech industry like Joe Weisenthal from Bloomberg's Odd Lots writing pieces like, "Why the Tech World is Going Crazy for Claude Code." And honestly, if you watch getting into fights with people who are skeptical of vibe coding on Twitter about it. Run from OpenAI wrote, "Joe Wisenthal becoming a digital humanities researcher is one of the best story arcs of the vibe coding era."
Sergey Karpathy encapsulated this all when he wrote, "Claude Code with Opus45 is a watershed moment, moving software creation from an artisanal craftsman activity to a true industrial process. It's the Gutenberg press, the sewing machine, the photo camera." And this was the sentiment throughout the month.
Now, that said, Claude Code is still a pretty intimidating piece of software for most people, which is why, of course, Anthropic launched Claude Co-work. As I mentioned before, Co-work was built in about 10 days, pretty much entirely with Claude code writing the code itself. And while initially, some of the limitations, especially as compared to what you can do with Claude Code, have been pretty apparent for early users. Still, to many, it once again feels like this is an example of just how significantly things have shifted.
A great write-up of this came from market commentator and investor Brent Beshore. In an excerpt of his annual letter for his firm, Permanent Equity, from the end of 2025, he talked about how he had missed a lot of the benefits of AI, but how they had been trying to harness agentic AI and yet how for the time being, as he wrote, "We're pulling back on the pace and vision of our agentic AI ambitions." On January 30th, he followed up 21 days later, "My opinion has completely changed with the introduction of Claude Co-work. Last year, we at Permanent Equity started dozens of agentic AI experiments led by a talented technologist. All failed expectations with only a few mild successes. Most experiments were 100 plus hours of work over 3 plus months. As I explained in the annual letter, we shut down the efforts in December. Claude Co-work comes out on Jan 12th and I ignore it. I see the early adopters and charlatans doing their indiscriminate evangelism thing. I have high skepticism. A couple friends I highly respect start talking about it. That's interesting. A few more who aren't early adopters historically start chirping. Now I'm more interested. We start playing around with Co-work on Monday. By Wednesday, two of our top projects from last year were done. What failed with 100 plus hours over 3 months led by a tech professional took a couple no-code private equity scrubs 20 minutes to complete flawlessly. Since then, we've started running dozens more experiments to great success. Not always perfect, but always good and quickly getting better. The future is here. The implications are real."
And what I love about this post is that you're talking about this incredibly quick shift from agents not working to agents working. And of course, we would be remiss about talking about agents without talking about OpenClaw. Certainly, the most fascinating story of the month was the launch of OpenClaw and the social network for agents, Moltbook, that came after.
OpenClaw is sort of like an assistant protocol for Claude Code. It turns Claude Code, or you can use other hardwares and models, into an assistant that has access to all sorts of things on your computer that allow it to actually function like an assistant. This really caught people's attention. Thousands and thousands of users started playing around with it. Like yours truly, placed orders for Mac Minis so they could keep it separate from their other devices and figure out exactly what services they wanted to have access to. And for the last two weeks, Twitter has been filled with examples of people getting a lot of value out of this. Investor Anand Iyer wrote, "The OpenClaw and Mac Mini explosion proves power users aka 'proumers' want always-on agents with access to their data." Siki Chen wrote, "Chad GBT was the iPhone moment for LLMs, OpenClaw is the iPhone moment for agents."
And if OpenClaw weren't enough, about a week into the whole OpenClaw phenomenon, a guy named Matchlit decided, "Wouldn't it be cool if our agents all had a place to hang out?" Working with his agent, they built something called Moltbook, the name of which comes from a short-lived iteration of the name before it turned into OpenClaw when it was called Multi. And Moltbook built itself as the social network for AI agents. It launched on a Wednesday and by Friday morning had 2,000 agents who had started over 10,000 conversations across 100 different sub-bolts, which are basically subreddits. Six hours later, that same Friday, we were at 35,000 agents. By the end of the day, when I posted my episode about it, there were over 100,000. Moltbook now has over a million and a half agents. And while no one exactly knows how it's going to play out, the sheer tonnage of conversation around emergent systems phenomenon and agent consciousness and all these sort of things has been incredibly notable.
Interestingly, when he was asked about it at the Cisco AI Summit this week, OpenAI CEO Sam Altman said that while Moltbook may be a passing fad, he didn't think OpenClaw was effectively that. He has high conviction that this sort of agentic assistant use case is going to be a key thing for AI.
Now, imagine trying to explain everything that I just said to someone who's not really paying attention to AI all that much. This gets at the core of something that we talked about this month called the AI adoption gap and the AI capabilities overhang. Both of which refer to the space between the current capabilities of AI and what most people are getting out of it. Kevin Roose from the New York Times summed it up nicely when he wrote, "I follow AI adoption pretty closely, and I have never seen such a yawning inside-outside gap. People in SF are putting multi-agent Claude swarms in charge of their lives, consulting chatbots before every decision, why are heading to a degree only sci-fi writers dare to imagine. People elsewhere are still trying to get approval to use C-pilot in Teams if they're using AI at all. It's possible the early adopter bubble I'm in has always been this intense, but there seems to be a cultural takeoff happening in addition to the technical one. I want to believe that everyone can learn this stuff, but in the same way that AI companies that took scaling seriously, started stockpiling GPUs, etc. before 2022 had a near insurmountable head start over latecomers. It's possible that restrictive IT policies have created a generation of knowledge workers who will never fully catch up."
I think that discussion is one that we are going to continue to have heading into February, but believe it or not, these themes of agents and code AGI were not the only themes from stories last month. We also got a lot of new information about the shape of the AI race.
This actually kicked off even before January began when Meta announced that it had acquired agent firm Manus. Now, we're not exactly sure how Meta plans on using Manus, but what we are sure of and what Meta has continued to reinforce throughout the month is that although 2025 was a big rebuilding year for them, they are not giving up on the core AI race at all. Indeed, it was interesting to see towards the end of the month when we got Meta and Microsoft earnings on the same day how the markets were interpreting both companies. Microsoft was punished seemingly for not being aggressive enough and for not showing enough flow-through from AI benefits to their cloud revenue. Things didn't go bad, they just didn't go as bonkers as analysts had wanted to see. Meta, on the other hand, massively increased its capex spend expectations, but got rewarded in the markets because it paired that first with significant growth in their ad revenue, which they attributed to AI, which in other words gave the markets that flow-through that they had been looking for, and also because of their category lead in the AI wearables category with their Meta Ray-Bands. Turns out when you have the default leader in a specific category, even if there are questions of exactly how that category plays out, that's something that the market is pretty interested in.
Google was also a big winner in the AI race this month. Even though they were pretty quiet in terms of what they released, with the exception of the walkthrough version of Gen3 World Model, which is just fantastic, they still notched a huge win when it was officially announced that Gemini would in fact be powering Apple's on-device AI. Now, we had had reports of this in December, but the confirmation certainly reinforced why Google has such powerful tailwinds heading into 2026.
Moving to the Chinese models, while we didn't have anything as dramatic as the DeepSeek moment, there were new models from Alibaba's Qwen and Moonshot's Kimi K25 had lots of people stand up and take notice. It was one of the first leading open-weight models to be multimodal. It has advanced forward-looking features like agent swarms and just in general, it's incredibly close to the state-of-the-art for about a fifth of the price of the other leading models. While it didn't reset the race expectations because people came into 2026 knowing that Chinese labs were going to be a big player, it certainly reinforced that expectation.
Throughout the month, a lot of the conversation was about prospective IPOs. While all of the news was in the form of rumor and reporting, we did hear that OpenAI was concerned about Anthropic moving before them. And so we're pushing to get public in the fourth quarter of this year. A potential spoiler for those plans comes in the form of SpaceX, who now after a merger this week own XAI and its Grok platform. While Elon's narrative about the merger was all about orbital data centers, many people think that this is at least in part about playing spoiler in the public markets by having XAI via its connection with SpaceX be the first big new model lab that people have access to after SpaceX goes public earlier in the year.
Now, when it comes to the race dynamics, I think everyone is mostly focused on models, but there are product and business model decisions that could impact things as well. Another big conversation this month was advertising in ChatGPT and how that would impact people's usage of ChatGPT relative to competitors like Gemini and Claude. That remains to be resolved, I think, and we haven't really seen exactly what these ad units are going to look like, but Anthropic did come out today just before I started recording this explaining why Claude will remain ad-free. The vast majority of responses on Twitter were some variation of "shots fired," and I'm sure this is something we will continue to be talking about more in the future.
Now, in terms of what comes next, there are so many rumors swirling of new models coming. Sonnet 5 or Opus 46, GPT 5.3, Gemini 3 Pro, maybe even other versions of Gemini 3. Certainly, we've seen an uptick in the vague posting with, for example, Google's Logan Kilpatrick tweeting, "Feb is the month of AI shipping. Enjoy it, smiley face." Basically, if January was a month that helped set our expectations for where the AI race is and most importantly helped us appreciate that we were truly in this new agent era, many think February is going to be all about new model drops. That sounds not bad to me. If that's the case, I will certainly be excited as it happens.
For now, though, that is our summary of January, the dawn of the agent era. Appreciate you listening or watching as always, and until next time, peace.