Building the future of agents with Claude

1. 0:00 – 0:30

   Introductions

   1. BABrad Abrams0:00

      'Cause as a developer, like, my creativity ends at some point.

   2. AAAlex Albert0:04

      Yeah.

   3. BABrad Abrams0:04

      I can only think of so many use cases.

   4. AAAlex Albert0:06

      Right.

   5. BABrad Abrams0:06

      But the model, like, anything, anything somebody comes with, the model will figure out a way-

   6. AAAlex Albert0:11

      Right

   7. BABrad Abrams0:11

      ... to go do that thing. [upbeat music]

   8. AAAlex Albert0:16

      Hey, I'm Alex. I lead Claude Relations here at Anthropic. Today we're talking about building the future of agents with Claude, and I'm joined by my colleagues.

   9. BABrad Abrams0:23

      I'm Brad. I run the PM team on the Claude Developer Platform.

   10. KLKatelyn Lesse0:27

       I'm Katelyn. I lead the Engineering team for the Claude Developer

2. 0:30 – 2:30

   What is the Claude Developer Platform?

   1. KLKatelyn Lesse0:30

      Platform.

   2. AAAlex Albert0:30

      Let's talk about the Claude Developer Platform. [chuckles]

   3. BABrad Abrams0:32

      Yeah, let's start with that.

   4. AAAlex Albert0:33

      Uh, let's start with that.

   5. KLKatelyn Lesse0:34

      Start there.

   6. AAAlex Albert0:34

      It used to be called the Anthropic API.

   7. BABrad Abrams0:37

      Yeah.

   8. AAAlex Albert0:37

      We just went through a big name change.

   9. BABrad Abrams0:39

      Yeah.

   10. AAAlex Albert0:39

       Can you walk me through why we made that change, and also what this new platform is and what it encompasses?

   11. KLKatelyn Lesse0:45

       Yeah, totally. So the Claude Developer Platform really encompasses our APIs, our SDKs, our documentation, all of our experiences within the console, and really everything that a developer needs to actually build on top of Claude. We're really humbled, proud to serve, um, some really awesome customers around the world who are trying to, what we like to say, raise the ceiling of intelligence-

   12. AAAlex Albert1:07

       Right

   13. KLKatelyn Lesse1:07

       ... um, using Claude. Um, and the platform really enables them to do that. Um, and I would say one of my favorite parts about it is the platform doesn't just serve customers, uh, externally, the platform actually serves our internal product.

   14. AAAlex Albert1:20

       Mm.

   15. KLKatelyn Lesse1:20

       So, um, we love telling people, like Claude Code, for example, is actually built directly on our public platform.

   16. AAAlex Albert1:27

       I see.

   17. BABrad Abrams1:28

       Yeah. I mean, uh, I think when we started, we were just the Anthropic API, and very simple access to the model. But over the last year or so, we've added so many features to it. Um, we added prompt caching, we added a whole separate batch of API, uh, we added web search, a web fetch, we have this context management support, the code execution. So all these tools-

   18. AAAlex Albert1:49

       Yeah.

   19. BABrad Abrams1:49

       Now, you know, now it's kind of we feel like, yeah, it's aspirationally where it's re- it's a platform now.

   20. AAAlex Albert1:54

       I see. So there's just a lot more to it now.

   21. BABrad Abrams1:56

       Yeah.

   22. AAAlex Albert1:56

       It's evolved in pretty drastic way over the past year, and-

   23. BABrad Abrams1:59

       Yeah. Yeah, I think so

   24. AAAlex Albert1:59

       ... better naming.

   25. BABrad Abrams2:00

       Yeah.

   26. AAAlex Albert2:01

       Um-

   27. BABrad Abrams2:01

       And I think that's what developers were sort of calling it anyway.

   28. AAAlex Albert2:04

       Yeah. Yeah.

   29. BABrad Abrams2:04

       You know, so it's always natural to just sort of go with what developers are saying.

   30. AAAlex Albert2:07

       Right. We were a little late to the game there.
3. 2:30 – 3:15

   What is an AI agent

   1. BABrad Abrams2:30

      Yeah. I mean, agents is... It's almost sort of a buzzword, right?

   2. AAAlex Albert2:33

      Yeah.

   3. BABrad Abrams2:33

      Like, everybody you talk to now is building agents, and, and whenever a industry tech term gets to that level, you know, the definition gets very gray, everything everybody builds is an agent. But Anthropic, what we really think about a agent is where the model is taking some, uh, autonomy to be able to choose what tools to call, to call those tools, to handle the results, and how to choose the next step. So, uh, as a, as a foundational research lab, like leaning into the model and be, and, and what it, its reasoning, how it decides what to do, we think that's a really important element of what an agent is.

   4. AAAlex Albert3:09

      Mm. So it's kind of like the, the aspect of it being autonomous in some sense.

   5. BABrad Abrams3:14

      Yeah.

4. 3:15 – 4:00

   Building frontier intelligence for AI agents

   1. BABrad Abrams3:15

      Yeah. Yeah.

   2. AAAlex Albert3:15

      Charting its own course.

   3. BABrad Abrams3:15

      I mean, I think there's also re- I mean, we have customers doing really useful workflows where they're sort of predefining the path that Claude-

   4. AAAlex Albert3:23

      Mm-hmm

   5. BABrad Abrams3:23

      ... should walk, and that, that is a super useful thing to do. But what's nice about the agentic thing is as the model gets better, every couple of months, you know, we release a new model, and it- with a true agentic pattern, you know, those services are just gonna get better. Where, where if you build a workflow with a lot of scaffolding in it, you kinda put bounds on the model-

   6. AAAlex Albert3:43

      Mm

   7. BABrad Abrams3:43

      ... which is maybe okay in some use cases, but that means that you're- you may not take advantage of the next level of intelligence that a next model release gets.

   8. AAAlex Albert3:52

      Yeah. So it seems like there's this interesting trend with agents, at least over the past 6 to 12 months-

   9. BABrad Abrams3:57

      Mm-hmm

   10. AAAlex Albert3:58

       ... where, like you've said, the scaffolding

5. 4:00 – 5:05

   Reducing model scaffolding to build better agents

   1. AAAlex Albert4:00

      has been-

   2. BABrad Abrams4:00

      Yeah

   3. AAAlex Albert4:00

      ... a bit of a hindrance, and maybe we're dropping some of that.

   4. BABrad Abrams4:03

      Mm-hmm.

   5. AAAlex Albert4:03

      Um, can you explain the intuitions behind that around is, is this actually the future? Is like we give less and less things to the model?

   6. BABrad Abrams4:12

      Yeah. I mean, I think over time what we're seeing is the scaffolding the model needs to be able to accomplish tasks is, is it's needing, it's needing less. As the, as the level of intelligence of the model goes up, and we have every... we believe is gonna keep going up-

   7. AAAlex Albert4:27

      Yeah

   8. BABrad Abrams4:27

      ... um, that basically the model has more contextual understanding of the high level task that it's trying to accomplish, so therefore it doesn't need as many sort of guardrails.

   9. AAAlex Albert4:38

      Right.

   10. BABrad Abrams4:38

       And in fact, those guardrails in some cases become some, uh, like a liability to have. Uh, we've had customers try out new models and say, "Oh, well, it's actually only just a little bit better." And then we kinda look into it with them about what's going on, and it turns out, well, yeah, they were constraining it in ways that makes it harder for them to see the intelligence of the model.

   11. AAAlex Albert4:59

       Ah. Does this match what we see in the field with, like, our customers where they're also following these

6. 5:05 – 6:40

   The evolution of agentic frameworks

   1. AAAlex Albert5:05

      same trends? I know at the limit we have customers exploring all sorts of innovative techniques for managing Claude.

   2. KLKatelyn Lesse5:11

      Yeah, totally. And there's actually a lot of, like, discourse about this right now, right?

   3. AAAlex Albert5:16

      Mm-hmm.

   4. KLKatelyn Lesse5:16

      Like, what is an agent and, and what does it need, what do you need to build? And, and there are people saying, you know, "It's just a while loop."

   5. AAAlex Albert5:22

      Yeah.

   6. BABrad Abrams5:22

      Right.

   7. KLKatelyn Lesse5:22

      Like, you don't have to try that hard.

   8. AAAlex Albert5:23

      Yeah. Right.

   9. KLKatelyn Lesse5:23

      And, um, I think ultimately, uh, there's a lot of, there's been a lot of evolution of frameworks that-

   10. AAAlex Albert5:30

       Mm

   11. KLKatelyn Lesse5:30

       ... people are putting around the model that are helping them orchestrate their agents, try to get the most out of the model. And, um, I think what, uh, the industry is maybe kind of circling around is a lot of that has become maybe too heavy-

   12. AAAlex Albert5:43

       Mm

   13. KLKatelyn Lesse5:43

       ... and maybe too opinionated, um, and which is why you kinda get the, the people coming back to, like, "It's just a while loop, and that is all you need." Um, and I think what we're trying to, to do there is to say-

   14. BABrad Abrams5:55

       Maybe in a lot of ways it is a while loop, but the things we can more uniquely do to help people get the most out of the model is a lot of those tools, those features, and otherwise.

   15. AAAlex Albert6:04

       Mm-hmm.

   16. BABrad Abrams6:04

       And so what we wanna do is put, um, you know, frameworks and tools and platform out there that is opinionated to some extent-

   17. AAAlex Albert6:12

       Mm-hmm

   18. BABrad Abrams6:12

       ... on how people should use those tools. Um, but it's not this like super heavy framework that really like, to Brad's point, gets in the way-

   19. AAAlex Albert6:20

       Mm-hmm

   20. BABrad Abrams6:20

       ... of what the model's ultimately trying to do. So to strike the right balance, it's like, you know, we've, we've seen what a lot of people have tried to do, so we know we can be opinionated there, um, but we wanna be lightweight [chuckles] in the way that we're doing that-

   21. AAAlex Albert6:31

       Right

   22. BABrad Abrams6:31

       ... and make sure that the real thing we're doing is helping you get the most out of the model, um, without, you know, bogging you down in some super heavy framework.

   23. AAAlex Albert6:38

       Right. So would you describe part

7. 6:40 – 8:35

   Unhobbling the model with tools like web fetch

   1. AAAlex Albert6:40

      of the strategy here then as providing these auxiliary tools and things that we can give to the model-

   2. BABrad Abrams6:47

      Mm-hmm

   3. AAAlex Albert6:47

      ... but we're not necessarily, like, placing the bumper-

   4. BABrad Abrams6:51

      Right

   5. AAAlex Albert6:51

      ... spawn, like the model itself or-

   6. BABrad Abrams6:52

      Yeah. We think about, we think about it as like, how do you unhobble the model?

   7. AAAlex Albert6:55

      Unhobble, yeah.

   8. BABrad Abrams6:55

      Like, the model already has a lot of capabilities, and in fact, I'm convinced that even if you take a current generation of models, there's way more intelligence in there than we've been able to unlock.

   9. AAAlex Albert7:04

      Mm.

   10. BABrad Abrams7:05

       But anyway, that intuition is if you just give the model, like, the tools it needs, and like set it, set it free-

   11. AAAlex Albert7:12

       Mm-hmm

   12. BABrad Abrams7:12

       ... let it be able to use those in the right way, um, you'll get great results.

   13. AAAlex Albert7:16

       Mm.

   14. BABrad Abrams7:16

       And, and I think like a good example of that is we launched this server side, uh, web search tool and we- and web fetch tools, and it's been interesting to watch customers use those. And, you know, all we did real... I mean, it's a very minimal prompt that we have. We just give it the web search tool, and, like, all of a sudden, deep research tasks are, like, almost completely done-

   15. AAAlex Albert7:39

       Mm

   16. BABrad Abrams7:39

       ... with, with just turning on that, uh, switch on the API because the model will call that tool, it'll look at its results, it'll say, consider it, and say, "Okay, maybe I need to call, you know, do these other searches."

   17. AAAlex Albert7:51

       Right.

   18. BABrad Abrams7:51

       And then, "Oh, that fourth link you returned, that's the great one."

   19. AAAlex Albert7:54

       Right.

   20. BABrad Abrams7:54

       It'll do a web fetch on that link and bring that data back, and really all that very autonomously on its own kind of deciding.

   21. AAAlex Albert8:01

       Right. I, I think it's almost kind of like an interesting shift in, like, where the intelligence of a system is being applied.

   22. BABrad Abrams8:07

       Exactly, yeah.

   23. AAAlex Albert8:07

       From like the developer having to apply their intelligence to guiding-

   24. BABrad Abrams8:10

       Right

   25. AAAlex Albert8:10

       ... towards like the model now-

   26. BABrad Abrams8:12

       Right

   27. AAAlex Albert8:13

       ... figuring it out.

   28. BABrad Abrams8:13

       And it's so exciting when the model does it because as a developer, like, my creativity ends at some point.

   29. AAAlex Albert8:19

       Yeah.

   30. BABrad Abrams8:19

       I can only think of so many use cases.
8. 8:35 – 10:50

   Building agents with the Claude Agent SDK (formerly the Claude Code SDK)

   1. AAAlex Albert8:35

      the developer platform, what do you recommend? What are some best practices or ways for me to get started?

   2. BABrad Abrams8:40

      Yeah. So, um, super tactically, actually, the number one thing that we recommend right now is the Claude Code SDK.

   3. AAAlex Albert8:47

      Mm-hmm.

   4. BABrad Abrams8:47

      Um-

   5. AAAlex Albert8:47

      Okay

   6. BABrad Abrams8:47

      ... and what's really, really interesting about the Claude Code SDK is we essentially built an agent harness, an agentic harness, um, around the model to run that loop, right, and automate a lot of that tool calling and otherwise feature use. And obviously, originally was built for coding purposes. Um, and what, uh, the team really quickly figured out was like, actually, this is like an excellent general purpose- [chuckles]

   7. AAAlex Albert9:10

      Mm-hmm

   8. BABrad Abrams9:10

      ... agentic harness. Um, and so what the SDK does is it gives people a perfect out-of-the-box solution to actually just start prototyping agents-

   9. AAAlex Albert9:19

      Mm

   10. BABrad Abrams9:20

       ... um, without having to go and build, you know, the loop with all the tool calling and otherwise. It's built on top of the messages API and all those same tools that, um, we're mentioning, but it kinda gives you that really great starting place right out of the box.

   11. AAAlex Albert9:33

       Right. I feel like this is a pretty common misconception, at least when I talk to developers, about the Claude Code SDK.

   12. BABrad Abrams9:38

       Mm-hmm.

   13. AAAlex Albert9:38

       So like, I'm not building a coding application.

   14. BABrad Abrams9:41

       Mm-hmm, mm-hmm.

   15. AAAlex Albert9:41

       Like, why would I wanna use this?

   16. BABrad Abrams9:43

       Yeah.

   17. AAAlex Albert9:43

       Mm-hmm. But you can kind of remove the coding-specific-

   18. BABrad Abrams9:47

       Yeah

   19. AAAlex Albert9:47

       ... parts, right?

   20. BABrad Abrams9:48

       Yeah. I mean, I think that's a great example of what we were talking about, removing scaffolding on the model. It's like once we got done removing things-

   21. AAAlex Albert9:55

       Mm-hmm

   22. BABrad Abrams9:55

       ... from the Claude, from Claude Code to really unhobble the model, it turns out there was [chuckles] nothing coding left.

   23. AAAlex Albert10:00

       Right.

   24. BABrad Abrams10:01

       When the, when you remove everything else, then it's just an agentic loop, and you're, you're really a minimalistic thing to give, uh, Claude access to, to a file system-

   25. AAAlex Albert10:11

       Yeah

   26. BABrad Abrams10:11

       ... to a set of like Linux command line tools, um, to the ability to, you know, write code and execute that code. So those are all very generic kind of-

   27. AAAlex Albert10:21

       Right

   28. BABrad Abrams10:21

       ... capabilities that turns out could solve a wide variety of problems.

   29. AAAlex Albert10:24

       Right. Yeah. I feel like something I've been running up to in like my own side projects and also seeing with-

   30. BABrad Abrams10:30

       Mm-hmm
9. 10:50 – 14:35

   Best practices for identifying agentic use cases

   1. AAAlex Albert10:50

      think that's super interesting.

   2. BABrad Abrams10:51

      Yeah. I mean, I think, I think the other really interesting to- thing to think about, especially for businesses looking at a- agents, is like, what use case to go target.

   3. AAAlex Albert10:59

      Right.

   4. BABrad Abrams11:00

      So, uh, thinking beyond the technology, like, what is the actual problem to go solve? And I, I, I think, you know, we've s- we see a lot of customers and like doing a lot of things. We love all of it, but where, you know, the biggest impacts are is where the customer has thought hard about what's the business value of this.

   5. AAAlex Albert11:17

      Mm.

   6. BABrad Abrams11:17

      Like, will it actually save this many engineering hours, or will it help us remove this much, uh, uh, manual work or whatnot? And being able to articulate, like, what you expect the outcome of the agent project to be, I think is really helpful in, in defining the scope, uh, of the agent.

   7. AAAlex Albert11:37

      Right. And t- tying back one more time to the SDK, so it seems like-

   8. BABrad Abrams11:41

      Yeah

   9. AAAlex Albert11:41

      ... it's been really, really useful for, like, individual developers like myself-

   10. BABrad Abrams11:45

       Yep, yep, yep

   11. AAAlex Albert11:45

       ... you know, starting out-

   12. BABrad Abrams11:46

       Yep

   13. AAAlex Albert11:47

       ... and just wanting to get hacking on something really fast. For these customers, for enterprises that are actually trying-

   14. BABrad Abrams11:52

       Right

   15. AAAlex Albert11:52

       ... to get real business value-

   16. BABrad Abrams11:54

       Right

   17. AAAlex Albert11:54

       ... out of these things, should they be using the SDK? Is it ready for them? Is it ready for scaled use like that?

   18. KLKatelyn Lesse12:01

       Yeah. So, um, I think in a lot of ways it is. In a lot of ways, if you are in a spot where you can... Like, you can deploy that runtime. Essentially, that's what you get out of the SDK-

   19. AAAlex Albert12:12

       Yeah.

   20. KLKatelyn Lesse12:12

       -is an agentic loop runtime. Um, you can go and deploy that runtime wherever you want, [chuckles] whenever you're ready to do so. But I think what we're really trying to do is take the spirit of what the SDK unlocks for people, like go kind of up to that higher order abstraction where we give you the loop, we give you a lot of the tool calling in an automated way, um, and say, "How can we learn from that and give people out-of-the-box solutions that, like, at scale-"

   21. AAAlex Albert12:36

       Mm-hmm.

   22. KLKatelyn Lesse12:36

       -um, will really be able to solve for their use cases?" And I think that's a lot of where we're kind of trying to go with our roadmap throughout the rest of the year. Um, and one really important bit when we think about that is if the entire goal here is to help our users, like, really raise that ceiling of intelligence, get the absolute best outcome out of the models, then higher order abstractions are not just make it easier because you don't have to write all that code yourself. It's actually, like, how can we, like, really truly help you get the best outcome-

   23. AAAlex Albert13:06

       Mm-hmm.

   24. KLKatelyn Lesse13:06

       -because we, uh, we're in the room with research, we're in the room with inference. Like, we know how to make sure that our abstractions, our agentic loop is going to be, uh, extremely powerful-

   25. AAAlex Albert13:16

       Right.

   26. KLKatelyn Lesse13:16

       -and extremely good at working with Claude. Um, and the last thing that I would add in there is especially as these things get longer running, and as we provide more and more tooling to help people get at those longer running tasks, um, another big problem that our users we know are gonna keep trying to solve is, uh, observability within-

   27. AAAlex Albert13:34

       Mm-hmm. Mm-hmm.

   28. KLKatelyn Lesse13:34

       -those longer running tasks. Um, and so that's, that's one of the most common things that comes up for folks is, you know, I, I have these long-running tasks, I'm trying to get, um, these really great outcomes. But, um, you know, I might need to do some steering, or I might need to tune my prompt, or I might need to think about tool calling a little differently. And, um, that's something that we know we can give people that observability through the platform over time-

   29. AAAlex Albert13:58

       Mm-hmm. Mm-hmm.

   30. KLKatelyn Lesse13:58

       -and that's another big area of focus for us.
10. 14:35 – 19:00

    Best practices for managing context and memory with Claude

    1. AAAlex Albert14:35

       of how-

    2. BABrad Abrams14:35

       Hmm.

    3. AAAlex Albert14:35

       -we're gonna address that. Um, before I do, uh, is there other tools that exist right now that folks should be aware of when they're getting started with the developer platform? Things have... you've found helpful or useful?

    4. BABrad Abrams14:48

       Yeah. I mean, I think there's a c- so we mentioned, uh, web search and web fetch. Uh, I think an-another big thing that we're seeing is, um, customers, m- uh, have to do... right now have to do a lot of work to manage the context window. So by default, Claude has 200k tokens of context. We have a mil-million token available now in beta on Sonnet, which is great, but even a million, there's a limit there. Uh, and what, what many customers have told us is that, um, they get better outputs, higher intelligence if they, uh, even use a smaller part of the context. And so we've done... We have a, a couple of cool features that are just coming out to help developers manage that context.

    5. AAAlex Albert15:31

       Hmm.

    6. BABrad Abrams15:31

       So in these agentic loops, a lot of times you're doing ten, 15, 100 tool calls, and you edit this file or look up data in this database or, uh, you know, send this email, and each of those tool calls takes up like a, you know, 100, 200, 1,000 tokens.

    7. AAAlex Albert15:49

       Right.

    8. BABrad Abrams15:49

       Uh, and so we have this cool feature that lets you, uh, lets the model actually remove some of the older tool calls that are not needed anymore.

    9. AAAlex Albert15:57

       Interesting.

    10. BABrad Abrams15:58

        Uh, and that give, that gives just, just like you, if you declutter your desk and declutter your notebook-

    11. AAAlex Albert16:03

        Yeah.

    12. BABrad Abrams16:03

        -like, you can focus a little bit better. So if you declutter the prompt, actually, the model can actually focus a little bit better.

    13. AAAlex Albert16:09

        Oh, interesting. So okay, we're removing unnecessary context.

    14. BABrad Abrams16:12

        Mm-hmm.

    15. AAAlex Albert16:12

        Is there a risk that we remove necessary context?

    16. BABrad Abrams16:15

        Yeah.

    17. AAAlex Albert16:15

        How does that work?

    18. BABrad Abrams16:16

        Yeah, yeah. Yeah, yeah. So, um, we have some, some guardrails and some-

    19. AAAlex Albert16:20

        Okay.

    20. BABrad Abrams16:20

        -bounds around it, so you don't... But g-the general rule is if you, um, remo- we try to remove the tools that are, like, several turns back.

    21. AAAlex Albert16:29

        Okay.

    22. BABrad Abrams16:29

        That the model's already made decisions based on those tools. But if you... AI was playing with it, uh, recently, and r-I removed the tools that it was just called-

    23. AAAlex Albert16:38

        Mm.

    24. BABrad Abrams16:39

        -and it's, "Oh, my tool results are gone. I don't know what to do."

    25. AAAlex Albert16:41

        Right.

    26. BABrad Abrams16:41

        And then the... but the mo- the Sonnet doesn't give up. Like, it's like, "I'm just gonna call this tool again."

    27. AAAlex Albert16:45

        Yeah.

    28. BABrad Abrams16:46

        You know?

    29. AAAlex Albert16:46

        Yeah, yeah, yeah.

    30. BABrad Abrams16:46

        Um, but yeah. So generally, we have put some bounds on that because of that experience. So we, we do preserve the most recent set of tools.
11. 19:00 – 22:10

    The future of the Claude Developer Platform (observability, computer use, and other ways to unhobble the model)

    1. AAAlex Albert19:00

       here. So it sounds like there's a ton of new features that we've recently launched. There's a lot of momentum, and now there's other offerings as well, like the Claude Code SDK-

    2. BABrad Abrams19:07

       Mm-hmm

    3. AAAlex Albert19:07

       ... and things coming out soon. Um, what are you most excited about, Katelyn? What's the, what's the future looking like here in the next six to twelve months?

    4. KLKatelyn Lesse19:15

       Yeah. So we talked a little bit about these higher orders of abstraction where we can, um, really just make it, uh, as, as simple as possible for you to get the absolute best outcomes out of Claude. Um, and we wanna pair that with the observability that we talked about, um, so that you can really like, you know, see the data and take those insights from those longer running tasks. Um, and if you combine these things together and start to think about some of the capabilities like memory that Brad just talked about, you can really start to see this flywheel where over time, we're not just able to help you get the best outcomes out of Claude, but we can help you get self-improving and continuously-

    5. BABrad Abrams19:50

       Mm. Mm-hmm

    6. KLKatelyn Lesse19:50

       ... improving outcomes out of Claude. And that to me is kind of the, the like galaxy brain magic of the roadmap, is get to a point where, um, you know, we, we have people coming to us, they're building on Claude, they're, they have their tasks, they know what they're trying to do, um, and they get these like really like aha moments where over time it's getting better and better and better. Um, and you know, that is... that's kind of the biggest thing that in everything that we're doing, we're trying to make sure we're going after.

    7. AAAlex Albert20:17

       That's awesome.

    8. BABrad Abrams20:19

       Yeah, I mean, I guess I'd have to say I'm a- I, I'm always excited about model launches.

    9. AAAlex Albert20:23

       Yeah.

    10. BABrad Abrams20:23

        Like, it's like Christmas, like what, how, what will, will it, what will be possible now? So I love playing with the model launches as they come out, just unlocks more use cases. Some use cases that, you know, we've been working hard on and, and trying to improve, which is satisfying to see, but also some things, oh, we had no idea the model would be able to do this thing.

    11. AAAlex Albert20:41

        Right.

    12. BABrad Abrams20:41

        You know, now it draws ASCII pictures so much better.

    13. AAAlex Albert20:45

        Yeah [laughs] .

    14. BABrad Abrams20:45

        Or what, you know, whatever.

    15. AAAlex Albert20:46

        The important things [laughs] .

    16. BABrad Abrams20:47

        The very important things. But beyond that, the other thing I'm really excited about is, um, we're, we're in the early stages of giving Claude a computer. You know, I think about if we, uh, hire an employee here at Anthropic, and we, we welcome them, "Here's your first day," but we don't give them a computer.

    17. AAAlex Albert21:03

        Yeah.

    18. BABrad Abrams21:03

        Like, they would not be very successful [laughs] at Anthropic.

    19. AAAlex Albert21:07

        Right.

    20. BABrad Abrams21:07

        So like right now, essentially everybody u- is using Claude and, and it doesn't have a computer. So I'm, I'm really excited about giving Claude a computer, and I... you see like the very baby steps of that-

    21. AAAlex Albert21:17

        Mm-hmm

    22. BABrad Abrams21:18

        ... with the code execution tool-

    23. AAAlex Albert21:19

        Mm-hmm

    24. BABrad Abrams21:20

        ... uh, where the, the model can write code, execute it on the VM, and get the results back. So like it can like zoom in on images or take a, a Excel spreadsheet and create like amazing data analysis with charts and graphs, and that's just the baby step. Like, what if it had a persistent computer-

    25. AAAlex Albert21:39

        Right

    26. BABrad Abrams21:39

        ... that was always there and it could like organize the files in there the way it needed-

    27. AAAlex Albert21:44

        Yeah

    28. BABrad Abrams21:44

        ... and get the tools set up the way it wanted and, um, I just think there's a lot of headroom-

    29. AAAlex Albert21:49

        Right

    30. BABrad Abrams21:49

        ... to, to that scenario.

Episode duration: 22:10