Claude Code 이후, Anthropic의 다음 히트작은 무엇이 될까?

Video Title: Anthropic's Hunt to Find the Next Claude Code

Video Author: ACCESS Podcast

Compiled by: Peggy, BlockBeats

Editor's Note: As large language models continue to advance rapidly and AI coding tools become more widespread, the industry conversation is shifting from "can the model complete the task" to "how can model capabilities be organized into products, workflows, and business systems."

Over the past year, products like Claude Code, Codex, and Co-work have entered developer and knowledge worker scenarios. AI is no longer just a chatbox answering questions; it's becoming a production interface capable of calling tools, executing tasks, and validating results. But as the consensus that "agents will be the next software paradigm" solidifies, a more critical question emerges: who can first transform model capabilities into reusable, distributable, and scalable work systems?

This article is compiled from an interview on the ACCESS Podcast with Mike Krieger. Mike Krieger, co-founder of Instagram, is currently the Chief Product Officer at Anthropic, leading Anthropic Labs. His mission is to guide the team in exploring Anthropic's next frontier product directions following Claude Code.

Alex Heath (left) and Mike Krieger (right)

In this conversation, Mike Krieger doesn't just speculate on Anthropic's next product. Instead, he deconstructs the AI product competition into a set of more fundamental structural questions: How do model capabilities integrate into real workflows? How should AI companies internally organize innovation? How do platform companies manage boundaries with ecosystem customers? And as AI execution becomes more powerful, where will human judgment be repositioned in the production chain?

First, the product paradigm is shifting from "chat" to "tasks." Previously, large models primarily existed as dialog boxes where users input prompts and models generate responses. Now, Claude Code, Co-work, and Claude Design represent a different product logic: enabling AI to continuously work towards a goal, calling tools, generating results, and performing verification along the way. This means the key to an AI product is no longer just response quality, but task decomposition, contextual continuity, tool invocation, and result validation capabilities. Whoever can package these capabilities into a seamless workflow will be closer to the next generation productivity gateway.

Second, the organizational method is shifting from "large team planning" to "small team experimentation." Anthropic Labs operates more like a startup unit embedded within a large company: starting with two or three people, reviewing progress every two weeks, and using high-frequency feedback to decide whether to continue a project. Previously, innovation labs in large companies often suffered from long cycles, ambiguous responsibilities, and projects that were "good enough" dragging on. Now, models have lowered the cost of building, making judgment, taste, and decision-making speed the truly scarce resources. This means organizational efficiency in the AI era depends not only on engineering headcount but on the ability to validate directions faster with smaller teams.

Third, the boundary between platform and application is being redrawn. The success of Claude Code means Anthropic is no longer just a model provider but is also beginning to define application forms itself. The controversy surrounding Claude Design and Figma shows that model companies entering the application space will inevitably touch the interests of their clients and ecosystem partners. Previously, foundational model companies primarily provided underlying capabilities, with vertical applications like Cursor and Figma handling the user interface and scenario packaging. Now, model companies also need to demonstrate the agent-first future paradigm through their own products. This means AI platform competition is not just about APIs but also about product paradigms.

Fourth, the stronger AI gets, the scarcer human judgment becomes. Mike repeatedly emphasizes that Claude can write code faster, generate prototypes, and execute tasks, but it cannot replace the hardest part of the 0-to-1 process: asking the right questions, understanding real users, defining the product's north star, and judging what is "right." Previously, execution ability was the main bottleneck in knowledge work. Now, execution is being accelerated by models, concentrating human value more on front-end judgment, creativity, relationship networks, and organizational ability. AI won't automatically eliminate difficult decisions; instead, it will amplify mistakes in the wrong direction faster.

If this conversation were condensed into one judgment, it would be: After Claude Code, what Anthropic is searching for is not a single blockbuster product, but a methodology to transform model capabilities into production systems. In this sense, the subject of this article is not just Anthropic's next product roadmap, but the structural turning point of the entire AI industry moving from a "model race" to a "systems race."

The following is the original content (edited for readability):

TL;DR

· AI product competition has shifted from "stronger models" to "how to implement capabilities," essentially meaning large model companies are competing for workflow access points.

· The significance of Claude Code is not just writing code; it proves that an agent can persistently execute tasks towards a clear goal, pushing AI from a chat tool to a production system.

· The core value of Anthropic Labs isn't the number of products launched, but in using small teams to quickly verify what capabilities the model should possess next.

· Co-work represents Anthropic's attempt to extend the Claude Code methodology to non-programmers, essentially abstracting "programming ability" into work automation capability for ordinary people.

· OpenAI Codex's pursuit means Claude's advantage is no longer just technological leadership; it depends on whether Anthropic can integrate Claude Code, Co-work, and Claude.ai into a unified experience.

· Model companies building applications themselves will intensify boundary conflicts with customers, but it's also an inevitable path for them to define the next generation of AI product forms.

· The faster AI can execute, the more human value concentrates on front-end judgment, product taste, and problem definition, because wrong directions will also be amplified faster by AI.

· The impact of AI on employment is not a problem a single company can solve. It will fundamentally force society to re-discuss skills reshaping, distribution mechanisms, and irreplaceable human capabilities.

Original Content

Alex Heath (Host): After Claude Code, what is Anthropic's next big product? On this week's show, we have Mike Krieger. He's the co-founder of Instagram and now leads the team working on "moonshot projects" inside Anthropic.

Mike Krieger (Chief Product Officer, Anthropic):

One of my darkest days at Anthropic was naming it 3.5 v2. I can explain why we ultimately chose that name.

Alex Heath: Mike and I recorded this conversation in person in San Francisco during Anthropic's recent Claude Code conference. At that conference, Anthropic announced a new large-scale compute partnership with Elon Musk. So, you guys are going to space with Elon now?

Mike Krieger: Absolutely. Yes, we are looking for new, even unexpected, sources of compute power.

Alex Heath: We talked about what Mike is working on now, the intense competition between Anthropic and OpenAI, and what parts of human work Mike thinks will still matter even as AI gets stronger.

This is Access.

Mike, great to see you in person at the Claude Code conference in San Francisco. I was just thinking back to our last conversation. You had just taken over Labs not too long ago, but it's been a few months, right?

Mike Krieger: Yeah, almost four months.

How Labs Operates: Bi-Weekly Triage, Validating Big Products with Small Teams

Alex Heath: Almost four months. For people who don't know about Labs, I want to start here. It's a pretty unique organizational structure. We talked about it a bit when I visited your office a few months ago. What exactly is Labs, and what is its mission within Anthropic?

Mike Krieger: Simply put, my understanding of Labs – I'd call this current version Labs v2. We can talk about what Labs v1 did and what Labs v2 wants to do later.

But I think Labs does two main things.

First, it's about narrowing the gap between Claude's theoretical capabilities and how ordinary people actually use it daily. That is, Claude can theoretically do a lot, but how do these abilities truly enter people's daily work and life? What products, prototypes, or projects do we need to make to show how to unlock more of this potential, to narrow this gap as much as possible?

Second, we act more like a "frontier reconnaissance team," figuring out which direction the model needs to evolve to meet the needs of different users.

So, a successful Labs project doesn't necessarily have to be released as a product. It could be a prototype. We build it and realize: the model isn't good enough yet to complete this task. So, we set it aside, re-evaluate it when the next-generation model is released, turn it into an evaluation metric for future model development, and then iterate.

Therefore, unlike product labs in pure product companies, where success might be measured by "did you launch a product," at Anthropic, the value of Labs can also manifest in other ways: it can influence Anthropic's future direction.

Alex Heath: Labs has made some hits, right? Claude Code is one. MCP is another. What else?

Mike Krieger: Agent Skills is another very important thing to come out of Labs. Also, I can talk about a project we didn't release but was hugely helpful for research: computer use, getting Claude to use a computer.

I joined Anthropic in May 2024. Next week marks my two-year anniversary, which we internally call "antiversary."

Alex Heath: Is it like an "ant-iversary"?

Mike Krieger: It's antiversary. Everything at Anthropic has to relate to ants. I was pretty resistant initially. We don't say dogfood; we say antfood.

After I joined, we started building Labs. One of the first projects proposed was: why not try to get Claude to use a computer?

Alex Heath: That's computer use.

Mike Krieger: Right.

Alex Heath: What model era was that?

Mike Krieger: That was Claude Sonnet 3.5. That was also the first model I was involved in launching. I started working on the launch in my third week. We often joke that Anthropic doesn't have a typical onboarding process; it just throws a really hard project at you. And for me, that was launching in my third week.

Sonnet 3.5 was an interesting model because it was one of the first to truly unlock some coding scenarios. Not full agentic coding yet, but you could see the beginnings.

So, we put Sonnet 3.5 in, built a computer use product around it. But it had many problems. For example, it was too slow at using the computer, accuracy wasn't high enough, and its vision wasn't good enough. It would see the screen, say "I need to click that button," and end up clicking somewhere else.

But building this "not fully functional" testing framework was incredibly helpful. Because later, when we got to Sonnet 3.5 v2 – the naming can be explained later; it was truly one of my darkest days at Anthropic – we could put the new model into this framework to test it.

We tried 3.6 later. It still wasn't good enough, but there was some improvement. By the time we got to 3.7, I remember that day vividly. I was on a business trip in New York, meeting the New York team. Suddenly, someone messaged me saying: we think that thing from Labs, the computer use project that had been sitting there for nine months, is starting to show real signs of life on Sonnet 3.7. We think it's time to open it up and publicly discuss computer use as a capability.

This was over a period of about nine months. Every few months, we'd put the new model into the same test framework. Even though Labs had put the project aside, it was still incredibly useful as a test suite for evaluating the evolution of computer use capabilities in models.

Alex Heath: When you first joined Anthropic, you were Chief Product Officer. I remember thinking at the time: Mike Krieger, the Instagram co-founder, a consumer product guy in my mind, why join an enterprise AI company?

Mike Krieger: Yeah.

Alex Heath: We might have talked about this then. I thought it was a really interesting choice. In hindsight, it was the right one. Of course, the timing was also great.

I'm curious. You initially joined as CPO overseeing the entire product line. The concept of an "AI product" is inherently a bit fuzzy and changes fast. How did you end up moving to Labs about 4-5 months ago? From what I understand, you're more of an IC now, an individual contributor? Do you still manage people?

Mike Krieger: I don't manage people. We're just about to enter the performance review cycle.

Alex Heath: So this is what you wanted, right? To escape writing performance reviews?

Mike Krieger: Exactly. I opened the system to see who I needed to review, and it turns out: you just need to write your own self-review and feedback for your manager.

Alex Heath: That's it?

Mike Krieger: That's it.

Alex Heath: Now Claude writes the performance reviews.

Mike Krieger: Claude does help write some reviews, which is useful. It won't write the whole thing, but it at least helps you remember: what did I actually do in the last six months?

I think the company goes through different phases, and how well those phases match what I'm truly passionate about varies.

When I first joined, the entire product and engineering team was maybe 30 people, perhaps split evenly. We certainly had engineering teams working on research infrastructure, scalability, etc., but if you look just at the people building products, it was mainly Claude.ai and what we then called the API – we didn't even call it the Claude Platform yet – totaling maybe 30 to 35 people, very very small.

It still felt very much like an early-stage startup. Many things were still being defined. Like, "what this product actually is" was far from settled. Claude.ai didn't have Projects, didn't have Artifacts. It was basically a list of conversations with Claude, with almost no extra features on top.

So joining Anthropic then felt a lot like joining a startup trying to find its product form. It already had tailwinds, of course.

Alex Heath: When you joined, the Claude 3 series was already out, right? Opus, Sonnet, and Haiku.

Mike Krieger: Yes. That was the first time Anthropic had a model series that was at least near the frontier. There was still so much to do on the product side: what is this product going to be?

Even though my background is more consumer-focused, the reason I was excited is that in the time between Instagram and Anthropic, I did a lot of investing with Kevin, my Instagram co-founder. We had a whole set of investment themes, one of which was "the future of work" – how will work get done in the future?

And Anthropic seemed very likely to unlock that theme: what happens when you have a very intelligent assistant to help you work? I didn't even foresee at the time how disruptive this would become.

Alex Heath: You probably thought: this is an interesting little AI company; maybe it can help me understand some investment themes.

Mike Krieger: Yeah, maybe it would help us understand the things we were thinking about. But actually, it changed far more than I imagined.

That was phase one: tiny team, projects you could count on one hand. Then fast forward to the end of last year, the product team had several hundred people. We had a whole portfolio of projects. A lot of work became about deployment, understanding customer needs, being customer-facing, management hierarchies, and all the things that naturally happen as a company grows.

I gradually realized that some people really love that kind of work and are great at it. I deeply respect them. But for me, I had a great coach who described this state as my "competence zone" – things you're good at, you do well, you can handle, but it's not ultimately what truly ignites you or drives you.

This is actually a dangerous position. You can stay in it for a long time, looking like you're performing well, but it's not where you have the most fire or motivation.

So, in Q4 of last year, I started discussing this with Daniela. I said the company has grown. We've essentially compressed the typical five-year growth trajectory. Though it's only been about two years.

Alex Heath: Yeah, I think you guys grew pretty well.

Mike Krieger: Yes, growth was okay. The team size and product portfolio expanded rapidly. So I said, I think I want to go build a new company.

Daniela asked: Is this because you want to leave Anthropic, or because you want to change what you do within the company? I said I really like this company. The people are great, I love the technology, the mission, etc.

Around that time, we were also restarting Labs. Because Labs v1 was too successful; all the projects "graduated," and eventually, no one was left. So Labs was essentially put away, set aside.

So we decided to restart Labs, and I got back into a builder role. Everyone seeing me inside or outside work says: "Mike, you look so happy."

Alex Heath: Some of your colleagues mentioned that to me earlier today. They said Mike is in such a great state, having so much fun.

Mike Krieger: Yeah. Of course, I'm still my harshest critic. So every day I think: how can I be doing better? What can we do