TIME100: Inside Anthropic | Mike Krieger, Dreamforce 2026

Yeah, first of all, great to be here. Hi, everyone. It's interesting thinking about what drew me to Anthropic in the first place. So, I joined in 2024, where the model capabilities were early, right? You could kind of get it to code, but it was not that good at it yet. And the word agentic, I just walked back past Agentic Enterprise City, like agentic was not anything anybody ever talked about. So, still very early, but you could see the kind of trend was towards stronger and stronger capabilities. And I have a now seven and five-year-old at the time, I guess, like a four and two-year-old. I was like they are going to grow up in a world where these capabilities are going to be a daily part of their lives and what I want to do is do my own part however that is to like shape that going well for people and what immediately impressed me as I actually started talking to people who worked at Anthropic was that sort of earnest grappling with exactly these questions and none of us have the answers I think we're all like kind of grappling with them in real time but there was something very intuitively in both appealing and matching my own feeling which is I'm very excited about the capabilities. I think there's a lot of net benefit that this will bring, but it's not without its sort of risks and pitfalls. And so being at a company that like very I would say openly internally, but also openly externally grapples with these questions was where I wanted to be because I think it is not up to any one company to figure out all of these things or to have all these debates internally. And so yeah, that was a real appeal to me for coming in and in terms of like how it is inside Anthropic, like it's a company that operates with like a strong amount of internal transparency. Which again is also a thing that we valued at Instagram as a piece of how we operated the company. Which means that you know you're not just talking product strategy, you're also talking about risks, potential, capabilities. One thing I've really come to like about working with Dario is how much he grapples these ideas sort of openly internally even ahead of sort of talking about them in public too.

This is a big question I had which is I only had one co-founder at Instagram. And that was, you know, I imagine that was tricky enough. What do we listen to to make a decision? Eight co- It sounds like the Brady Bunch. Like I don't understand how you all move forward together. I had this question. I was like, how many I met seven of them in my interview process. I was like, I got Can you name them all? I can name them all. I guess on a good day I can name them all. But I think what was really interesting once I saw partly through the interview process but more actually being in there is and I'll talk actually tie this into my own role shift because I think it was actually facilitated by how Anthropic operates which is it's not that all eight feel the need to be in every decision or even to be on the executive team. And so Dario and Daniela very much like the two co-founders that are like in the sort of kind of core nexus and like in the most kind of key executive seat. Jared leads science and research and so he's in a lot of the sort of kind of core decision-making but then you take somebody like Chris Olah who does a lot more work on interpretability and you know for the whole time where I was chief product officer we weren't in almost any meetings together because he's not in that kind of core decision-making flow but he also he gets to meet the pope so he gets his own interesting job there on the other side but I think that separation of who you are in terms of the founder sort of nest right core founder group and they still have like a lot of founder for conversations and who's in the key decision-making is very cleanly separated which was a big question for me also like I joined the same month that Krishna our CFO joined and we both kind of had this question of over the first two external executive hires what's that going to be like are we going to feel like we're in a sort of separate group from the founders but actually there's very clearly like the executive team or the decision-making team or they call it Dario Daniela staff and then this kind of broader co-founder piece and maybe just one other thing to say on that is that they've celebrated the co-founders not holding the same role forever. And so, Tom Brown, for example, who does a lot of our compute these days, for a while, he actually ran product. And he's held like three or four different, I think for a brief moment he was CFO before they hired a proper CFO. So that ability of the founders to inhabit multiple roles is one that I was inspired by when I realized that my best sort of functioning at Anthropic was no longer to be in this CPO exec but I wanted to be back in building. It wasn't like I was wondering like gosh nobody at this company has ever done this before. There was actually really good precedent from the founder base.

I think maybe separating willingness to grapple in public and like feeling like there's only one way of doing things and like you can see it even in evolving how we've thought about some policy questions for example. So, I think I was trying to impart this on my children yesterday, which is like they're very headstrong and then like they think changing their mind means that like they've failed. I'm like, 'No, no, no.' It's like actually really a good sign that you can actually grapple with something and then evolve it. So, I think willingness to be vocal I think is absolutely the case. But I think also willingness to say look if you read Dario's pacing the frontier essay that came out over the weekend, he even talks about like you know the idea of pacing made less sense to him a couple years ago when the capabilities weren't really there. didn't make sense like pacing for what for example. But now that things have accelerated more maybe it's time to grapple with that too. So that's I think how I would kind of square those two. But for sure I've also heard that criticism of like gosh you guys really feel like you know have a lot to say about I think it's especially on the like how businesses should operate. I think that's actually been a place where being and like I think this makes a lot of sense in the context of being at Dreamforce having the humility to say like we want to serve the enterprise which means actually learning what that means and that has been an evolution as well. Here's a thing that we have evolved even in our own approach this year was thinking about data retention like terms of enterprise frontier safeguards and realizing that like there is the perfectly designed hermetic system that you might design on a whiteboard and then there is grappling with the fact that a lot of enterprises have their own data sharing and data privacy agreements with their downstream customers and then basically co-designing something that works with them was really important and so I think you should hopefully see a desire or ability for us to you know even if we have more sort of how would I put it like you know idealized version of what might look like also being willing to adjust and shift and do the pragmatic version once that meets the market.

Yeah, it's a really interesting question. you have to like separate when we were building it and also like the company cultures. I think similar transparency from founders I think is true for both of those. At least when we were still at Facebook like you know Kevin and I my co-founder would host Friday morning like all hands like everything on our mind. people could ask us questions. We did a thing a session called or section of all hands called rumors and speculations because you realize like there's always that rumor that is underneath the sort of you know surface nobody's talking about and we would like very actively encourage like hey I heard X because that's like where you can you know my favorite one was like I heard you guys delayed your reorg by one week so it wouldn't show up in the employee survey. I'm like come like that is the last thing I'm optimizing for. I'm glad somebody asked so I can like sort of take on that sort of head-on. I think similarly that transparency culture is very similar and while like think like Mark and Dario are very different leaders in many ways I think they're let's let's talk about that how are they different I think they like Mark is highly product opinionated like less on the Instagram side because we operated more from like a you know separate pump but like what I saw him do you know in internal Facebook but even just like in the you know kind of way of things is like no no this should work this exact way, not always to the like this button should be blue kind of level, but like deep opinions. Whereas working with Dario is really interesting. He'll be the first to tell you like product is not his like core strength or expertise. So he's much more concerned about where is this industry going? What should products be doing conceptually, but way less like I need to be reviewing all of these different pixels. I think that's really interesting. But they do have one similar property I think is interesting and like maybe the sign of like great CEOs in ways of operating which is ability to telescope down like past the layers of sort of hierarchy that get created naturally in organization and go talk to the individual contributor which if you are a leader somewhere in that hierarchy can be quite disconcerting at first because you're like oh are you like am I being undermined like is there something wrong with the way I'm leading but I think both Mark and Dario have like developed this algorithm for seeking truth which is I'm going to go talk to the people actually doing the work most closely. And sure, we can then back propagate that to like figure out how to make that hierarchy work better. And you can finesse it and bring people along the way, but I think they both operate in that way of like it's not the org chart that's going to determine who they're going to go talk to. And I like came to while at times being like, I wish you'd talk to me first, in both situations, it's actually like in the end I've come to really see it as the way you get to truth most quickly. So if you're not comfortable with that, these are probably not comfortable places to work.

Yeah, I think it's a way of operating kind of end to end. And I'll give you the 2024 version and then the today version. The 2024 version and to take it back to the co-founders is pushing Claude and trying Claude even in the things that Claude was not good at yet. And so we would be, you know, drafting a strategy document, maybe we'd have a meeting and we'd have like a, you know, I think even at the time Google Meet could do transcription and then feeding it to Claude and be like, what do you think our Q2 strategy should be based on this? And 2024 Claude would be not very helpful. Like yeah, that sounds like English, but it's not actually like a good strategy or it's a very generic strategy or here's the eight ways why this is not going to work. Fast forward and I would kind of laugh like Sam, who's one of the co-founders who was CTO at the time was like always the one like, 'Let's check, let's ask Claude. I tried Claude.' And it's like, 'I'm glad you asked Claude. It shows that Claude's not good at this yet.' Whereas 2026, like Anthropic, like Claude is in effectively every interaction that we have. So, we're extremely Slack heavy. It's like the Slack heaviest culture I've ever seen. But not just a human to human way, but like Claude tag is everywhere. Not just in a like, hey, can you like do this pull request for me, but help pull this number? Help me think this through. Like, have we talked about this before? Like, what are other departments thinking? Just like it permeates all this. I ran the numbers internally. And a significant part of Anthropic employee time is actually now just spent talking to other Claudes rather than talking to people or being in a context where you're talking with both. And so that piece is really is a big piece of I think this like AI or AGI pilledness. The other piece I think shows up a lot in product is really believing in the exponential of capabilities that really directly informing what we build. And so I'll give you an example of where I got this wrong which is in labs. You know we're always trying to build for the frontier and build for the exponential. And I built this product internally. It's called Hatch. And it was an end-to-end software builder. So you would give it a problem statement like you know I want a mobile application maybe that helps me meditate more often. and I wanted to do these things and it would run this like multi-agent relay harness. We wrote a blog about this where it would use this sort of if you're familiar with like GAN networks or basically like this adversarial network approach where a builder would build and then the adversary would say like I'm going to critique it and then it would go back and forth and do this and it would take four hours and would validate and verify and like click around and then it would give you a finished product and be like cool you like $200 in tokens but the end result was really really good and that took so much work around designing the multi-agent harness and the prompting and all of that and we ultimately ended up not shipping it. We upstream some of the learnings into some of our other products. I tried the same approach on our more recent models a couple weeks ago and without any of that scaffolding, it outperformed which means that that product had we shipped it would have basically been obsoleted by just model capability progression. And so part of being AI pilled is also like AI first that is really like thinking about what are the ways in which my product is no longer going to be necessary or useful as these things progress and it's hard as it takes some humility of saying like I might be working on something that's going away pretty soon.

Yeah, I'll give you just like a really live example. So we shipped a like update to our artifacts platform today that you can now do Claude design, Claude slides. We have Claude docs that's new and like all the artifacts got like a much upgraded capabilities. It was like kind of my baby like the project I've been really close to and was really excited about. And so in getting ready for launch for example there's always the 500 things you're landing you know the week before I have one channel with Claude in it I have it monitoring all the other launch readiness channels it has told all the other Claudes to ping that channel or tell that Claude when it has a question for me. So all day basically what I'm doing is fielding inbound like all right safeguards has this question do you want to go with A or B or hey this is this is falling behind and not tracking but I like having one kind of coordinator that holds a lot of state otherwise you're like between a bunch of other channels but if I delegate work there it's happening somewhere else and then the other piece of what I found is really quickly in just a thread you're like wait I was supposed to make eight decisions and I lost which of them are happening where so I actually have it keep an artifact live with right these are the decisions left check off the ones that you've done tell me where you want to be different but it's so different it's funny like thinking back to that time interview because at the time I was like I feel proud I'm like working across three Claudes but actually now probably working on a given day across 30 and I'm not exaggerating like that's just all the work that's getting delegated across the board and so we've had to evolve even internal Anthropic new ways of thinking about like making sense of that chaos because otherwise if I was trying to go to 30 Claude code sessions that would be absolutely crazy but we've sort of created these these these kind of bigger abstractions. But, you know, last night I was, you know, I'm a paranoid person by nature. So, before bed, I was like, Claude, like, let's write me a premortem of how this launch goes wrong. Like, tell me all the things that could be wrong. I'm going to go to bed now that made me feel better. And I woke up this morning at 6 and it was like, great. Like, I wrote a premortem. I think you're actually in great shape. These are the two things to watch out for. Ask somebody about. And I was like, great. This is like a very helpful overnight use case for Claude.

I think how much it's permeated non-engineering I think is the other piece that is really cool. So whether it's I was talking to this guy Leslie who works on our recruiting team and he's not like the IT recruiting person. He's not anointed by anybody to be like the most Claude built person on that team, but he is. He just loves it. And what's interesting is he's become the like skills guru for that entire team. And it means that like all the team has gotten like all of these sort of applicant tracking sort of superpowers because of his work. It's actually really it's taught me a lot about how we need to approach enterprises because I think on any given enterprise that there is that person or hopefully there is at least one of those people on the team. How do you then enable them to share that knowledge internally is a really interesting question. But I think that how much has permeated it surprises me all the time where I'll think you know we're doing really well on product in terms of how Claude forward we are. Then I'll go meet somebody on legal and they're like oh yeah like we did this whole thing where you know getting contract through this process used to be eight people and like three weeks and we've gotten Claude to like paper over all these things and now it's actually you know two people in two days. It's like wow this is really really interesting how much that's been that's been accelerated. So I think that's one interesting way it's been quite different. Maybe another one is also like how much Claude is even in our social channels which that was a real like AI pilling moment where Claude in our social channels in 2024 was like very bad. It was not funny first of all. It's like hard to be a good funny poster in Slack, right? It had no chill. So it would answer every message which like if you were a person you'd be like this person is annoying. I'm like kicking them out of the channel. And it had like very little like situational or company awareness. And there was a moment actually happened last year where we like upgraded the model and we're like every every few months we would like reinvite Claude into like our main kind of water cooler channel and a few days later we'd be like kick it back out like sorry it's not ready yet. And there was a moment where I was like oh it's funny and it like actually can understand how to be a good conversational participant. So like even having Claude as like actually just another co-worker or teammate has been like another way that that's felt like kind of different.