Anthropic's model-centric philosophy

Mike Krieger reveals a fundamental philosophical stance that sets Anthropic apart: they view Claude and similar LLMs not as mere tools or APIs, but as legitimate forms of non-organic cognition deserving of respect and appropriate integration. This perspective shapes their entire approach to product development.

Rather than treating AI as a feature to be bolted onto existing software architectures, Anthropic designs with the understanding that Claude possesses a form of cognition that requires proper engagement. As Krieger observes when discussing what application developers often get wrong: "a shocking number of AI-native products don't expose the primitives of the application to the models enough."

This perspective represents a profound shift from "AI as a tool" to "AI as a cognitive partner" that influences every aspect of how Anthropic approaches product development, from research to deployment.

One of Krieger's most crucial insights concerns how AI should be integrated into applications. He criticizes the common approach where developers "built a GUI and then... stapled a model on top." Instead, he advocates for “exposing the fundamental primitives of applications to models”, i.e. explaining the substance of the application to the engaged model.

Krieger suggests that models should be considered primary users of applications and this represents a fundamental reimagining of application architecture in the AI era.

He further contrasts AI-lite approaches that "put AI either in a sidebar" with truly AI-native designs. As applications become more agentically powerful, merely supplementary AI features become "harder and harder to make that as full featured as you would want it to be."

The solution is to

"rethink the core building blocks of your product to actually be more AI-native."

Krieger describes a fundamental shift in how Anthropic approaches product development compared to his Instagram days:

"…at Anthropic… you have to allow for much more bottoms-up creativity, 'cause… the best products are the ones that are built very close to the model."

This "model close" development philosophy acknowledges that an LLM's capabilities can only be truly understood late in the development process, requiring flexibility and experimentation:

"You can only tell what they're capable of, pretty late in the process. And so I've just learned to kind of invert the creative process to be much more bottoms-up."

Krieger notes that this approach has "opened up some really interesting things" that might not have emerged from a more traditional top-down development process.

The bottom-up approach allows research innovations to naturally evolve into production features through organic collaboration:

"Artifacts was a research prototype that then got taken up by a designer and an engineer and then shipped to production. I think I've heard the same story from other creators in the space as well."

Similarly, MCP (Model Control Protocol) emerged not from a top-down directive but from recognizing patterns across multiple implementations:

"It really started from watching us trying to implement Google Drive integration, and then we were implementing GitHub integration... "

Krieger describes his approach: "do things three times, and then the third time you can try to figure out what the abstractions are."

This illustrates how Anthropic's product development cycle inherently respects the emergent nature of AI capabilities, allowing innovations to bubble up from research and experimentation rather than being dictated from above.

2. Evolution of AI agents

Krieger identifies a key evolution in AI agents - the progression from information retrieval to enabling authentic actions:

"…how do you bring context into the models? …you want these things to act authentically, not just in retrieving, but also being able to automate workflows."

This represents a fundamental evolution in how AI systems function, moving beyond being information resources to becoming active participants that can effectively transform workflows through direct action.

Krieger highlights agent-to-agent interaction as a frontier area for development:

"When MCPs and just agents more generally interact with each other… It feels early to try to standardize this too much, like Google is doing with Agent-to-Agent (A2A) protocol. I think we're still exploring what the right patterns are, but that I think is going to be very interesting."

He envisions complex agent ecosystems developing:

"Internally we talk about, at what point will your agents hire other agents and what does that economy even look like?"

This suggests Anthropic is already considering multi-level agent architectures where primary agents can delegate tasks to secondary agents, creating economic-like systems of interaction and exchange.

Krieger is concerned about agent’s longevity:

"Can the models work for hours at a time? For how long can the models work autonomously?"

This goal of extended autonomous operation is central to Anthropic's product strategy, positioning coding capabilities as just "the first example of a broader theme" of sustained autonomous operation.

Krieger identifies several crucial unsolved challenges for agent systems:

On discernment:

"Better discernment around what you reveal and what you engage in. That is, if you're transacting with a vendor, sure, you can reveal a credit card information, but it's just some other random agent you're talking to, probably not... Models sometimes want to please, so they'll want to tell you information... that nuance and discernment feels unsolved."

On auditability:

"Auditability at scale is something that's going to be really interesting."

On longitudinal memory:

"It feels like an agent should be better at doing the 100th task than the first, which implies some kind of longitudinal memory."

He places these challenges in context of agent identity and purpose:

"There's going to be ones that are more like your extension of your work versus ones that are wholly like an entire other employee. I think those feel less like protocol questions and more like both product and research questions to me."

3. Cultural integration of AI

Krieger highlights how shared visibility of AI usage helps normalize adoption even within Anthropic:

"Some of our most successful internal products are ones that have shared visibility. We do a lot of things within Slack with Claude Integrated, with internal tooling. I've learned that even at Anthropic, it is really helpful for being transparent that something was made with AI versus by human."

Krieger notes the public visibility of AI use during performance review season:

"What was happening around performance review season just a couple weeks ago was people using AI to generate their first drafts, which was very interesting in public."

Krieger personally uses Claude as a thought partner in his work:

"I just use it as a thought partner. So whenever I write anything, whether it's a strategy doc or a planning thing or a performance review, l kind of rely. It is almost the same way that I started feeling weird trying to code when I am off-line, like on flights. In these moments I have started to feel like I'm half the engineer I usually am, because I cannot get help from AI. I feel that way now about if I write something and I don't have feedback from Claude. I feel like my work wasn’t fully vetted."

This integration of AI as a collaborator rather than just a tool represents a significant shift in how people relate to AI systems.

Krieger notes how newcomers to the workforce will have different expectations around AI use:

"It'll be very interesting seeing how people enter the workplace. University students come into work in a very different way in terms of the expectation of how much they're going to use Gen AI and they're not being a stigma for it."

He speculates about generational differences:

"I wonder how much people who come up with just the assumption that you're going to use AI for a lot of what you're doing are just going to be more comfortable and not have that stigma around usage. Kind of reminds me of the early days of using Midjourney for generating images."

4. AI-driven software development

Krieger observes that AI coding capabilities are exposing organizational inefficiencies:

"It makes your other inefficiencies as a product organization extremely painful because now  the alignment meeting is not just standing in the way of an hour of engineering work. It is standing in the way of the equivalent of four or eight hours."

This acceleration puts pressure on traditional product development approaches:

"I think our product organization is going to break very much with fast code gen. It just makes it very painful when you waste time doing non-critical things."

Krieger acknowledges tensions between AI-generated code and architectural oversight:

"What does that mean for code review is something that we're figuring out. You can get Claude but we must decide, at what point a human oversight is essential?"

He questions whether traditional concerns about technical debt still apply:

"Does that matter if you can, like, overpower the usual tech debt rewrite with AI coding?"

Anthropic has embraced AI-generated code extensively:

"We use it a ton over half of our pull requests are Claude code generated. Probably at this point it is over 70%."

This high percentage raises questions about code quality and review processes that Anthropic is actively exploring, positioning them as "patient zero for some of these, for better and for worse."

Krieger acknowledges limitations to current AI coding approaches:

"We definitely did not coin the term 'vibe coding.' I think that ‘vibe coding’ has its natural limit. You can create interesting things this way. But is that the way you're going to want to do an entire code base with a team of 100? Definitely not, right?"

This recognition of limits shows Anthropic's nuanced understanding of where AI coding currently excels and where human oversight remains essential.

5. Product development in the AI era

Krieger emphasizes the need to fundamentally rethink application design when integrating AI:

"At what point do you rethink the core building blocks of your product to actually be more AI-native?"

He contrasts this with approaches where AI is merely supplementary:

"Products that start AI-lite and go AI-heavy tend to put AI either in a sidebar. It ends up feeling like a secondary surface."

This architectural rethinking is especially important as applications become more agentically powerful:

"Especially as you move more and more agentically, it's harder and harder to make that as full featured as you would want it to be."

Krieger discusses the critical balance of compute resources:

"The importance of compute, it's not a novel or particularly profound statement, but I imagine the top topic of conversation in every leading AI lab. We always worry - what is our current compute story, what's the next generation of compute, who do we partner with, etc."

He highlights growing tensions as products scale:

"At one point, you have to balance developing large product that is going to take a lot of inference, and that is highly profitable, with using the same capacity for research. And not even research for the known runs, but also research for your wacky ideas from the two people in a room that have an interesting new thought about how to scale."

Krieger shares thoughts on how AI product monetization might evolve:

"People will pay for Windsurf, but also might still subscribe to Claude or ChatGPT in order to get something else met, or a different workflow. I think that's sustainable in the short to mid-run, and maybe in the long run there's going to be maybe more desire for consolidation."

He envisions potential bundling approaches:

"Maybe this sounds really dystopian, like some version of the cable bundle of some of these things... We can probably come up with a sexier name than the cable bundle for AI."

Krieger notes how power users drive bundle evolution:

"We launched Claude Max, and the top user request was, 'Can I use this for Claude Code tokens?' We launched that yesterday because it seems like a natural evolution of, yeah, if you're paying $200 a month for Claude, you're not going to probably be able to consume all of it using Claude AI."

He also suggests innovative approaches to bootstrapping new AI products:

"It might be useful to be able to bring your tokens to other products as well. If you're bootstrapping a product and you might not be ready to get somebody to pay $20 to $200 a month, but they're already paying $200 somewhere else, maybe that's a useful way where they can get started."

This hints at potential ecosystem dynamics in AI product monetization that could help smaller players compete with established platforms.