Every Deploys Anthropic Agents in Slack, Saying AI Expands Work Volume

Every在Slack部署Anthropic智能体，称AI反倒增加工作总量 部署自主智能体会增加而非消除企业工作，因为基础任务的自动化拓展了项目野心，并将员工转变为系统指挥者。 Dan Shipper与Willie Williams详细介绍了Every如何利用Claude Managed Agents，从零散的个人机器人和脆弱的本地Mac mini过渡到直接嵌入Slack的全公司统一AI同事。他们解释了为何标准学术基准无法评估实际的知识工作，并展示了如何将人类品味转化为自动化测试和可复利的技能，例如将主编的编辑标准扩展应用到30,000处历史修改记录中。讨论涵盖了关键技术挑战，包括沙盒化预发布环境、上下文隔离、智能体社交礼仪以及通过computer use实现的桌面自动化。最终，他们指出，部署自主智能体并没有通过自动化消灭知识工作，而是通过将人类转变为高维度的系统统筹者，拓展了团队能够达成的能力边界。 Every基于Claude Managed Agents构建了其中央内部协作伙伴Every Agent，以分担执行沙盒、持久化记忆和基础设施维护的负担。 (12:59) 通过将30,000处历史编辑修改提炼为名为Kate Bench的自动化技能，Every在发稿量增长十倍的情况下，规模化延展了其主编的审校品味。 (6:26) 将智能体部署至共享Slack频道实现了公开提示（public prompting），消除了复制粘贴的繁琐，并将原本孤立的高级用户工作流转变为在组织内广泛传播的习惯。 (2:47) 针对自动化焦虑，Every发现积极引入智能体会扩大知识工作的范畴和抱负，而非削减就业岗位，从而在更高维度上创造了更多工作。 (30:51)

Dan Shipper and Willie Williams of Every argue that knowledge work is transitioning from manual execution to agent orchestration, where the central challenge is codifying human judgment and organizational context rather than pursuing generic benchmarks. Centralizing AI Agents in Shared Workspaces (2:47) Every initially experimented with decentralized setups running on dedicated Mac minis and individual one-to-one bots, but found that maintenance overhead was unsustainable and operational knowledge remained trapped in personal silos. Consolidating these tools into a single shared agent within Slack enabled public prompting, allowing team members to observe agent interactions in shared channels and adopt effective prompting practices organically. Codifying Taste into Compounding Skills (6:26) Shipper and Williams liken public AI benchmarks to SAT scores, arguing that generic academic metrics fail to measure whether a model aligns with specific domain standards. To bridge this gap, Every converts subjective editorial preferences into repeatable unit tests by capturing staff reactions to model outputs. To scale editorial operations without overworking staff, Every analyzed 30,000 historical edits from Editor-in-Chief Kate Lee to build 'Kate Bench.' The agent applies mechanical and stylistic revisions directly within Google Docs, while the team logs discrepancies between agent suggestions and final human edits to continually refine the skill. Offloading Infrastructure to Managed Execution Platforms (12:59) Managing custom infrastructure, isolated execution sandboxes, and persistent memory internally created significant operational drag. Every transitioned to Claude Managed Agents to offload low-level server orchestration, enabling their engineers to concentrate on agent interaction design and workflow integration. This architecture allows developers to clone production memory into isolated staging sandboxes for cost and safety stress-testing. Tight integration between the model and runtime harness also reduced model upgrades to a single line of configuration while enforcing granular permission boundaries based on user identity. Social Interaction Norms and Expanding Workloads (23:15) Moving agents into shared team channels creates distinct interaction challenges, as models must learn social etiquette—such as determining whether to speak or remain passive, adjusting personality parameters between assertiveness and patience, and preserving brand voice across complex message threads. Shipper disputes forecasts that automation will eliminate knowledge jobs, observing that automating baseline tasks expands organizational ambitions and increases total workload, shifting workers from direct task execution to high-level system orchestration. While agentic systems can successfully scale specialized human taste and eliminate redundant operational friction, their practical deployment remains dependent on frontier providers addressing subtle user-experience factors—including social conversational norms, personality configuration, and high-fidelity alignment—rather than raw benchmark intelligence alone.

[2:18] Dan Shipper: the only subscription you need to stay at the edge of AI. [4:33] Dan Shipper: I can't believe you're 35. I can't believe I have to still explain this to you. [9:25] Dan Shipper: How good is it at the work I do? [12:59] Willie Williams: Infrastructure is hard. [12:59] Dan Shipper: You're like, 'Oh, infra is not that hard anymore. We'll just have Claude spin up a bunch of servers and manage the servers for us and blah, blah, blah.' I was saying that, at least, and Willie was like, 'It's hard.' And I was like, 'We have to try it.' And then a month later, I was like, yeah—this is hard. [12:59] Katelyn Lesse: Thank you for your service [18:42] Willie Williams: acting as close to a human as it can within Slack [19:33] Willie Williams: We take humor seriously [23:15] Dan Shipper: use Hands to do it on my computer. [30:51] Dan Shipper: The more we use agents, the more work there is to do. [32:02] Willie Williams: Remember when Dan hooked you up? Have you claimed that beer yet?

[0:06] Super excited to be chatting with you guys. [0:08] I have said this to you before, but I think you guys are some of the most [0:11] AGI-pilled people out there with the things that you build and the way that [0:14] you talk about what's happening in AI. [0:16] So very exciting for me to get to spend some time chatting with you guys about [0:19] what's going on and what you're building. [0:21] Maybe, actually, just to start, I would love to talk about the fact that [0:25] the way everybody is doing work is changing so much right now. [0:29] That means the tools that people need to be able to use to get work [0:32] done are changing a lot. [0:34] How do you guys think about that? [0:36] We think about it a lot. [0:38] It's one of those questions that it's on everybody's mind. [0:44] Everybody is super afraid of it, or they're like, "Oh, [0:47] it's going to be a utopia, and no one's going to have to work at [0:49] all." The space we try to occupy is, well, why don't we just use the [0:52] models and try to figure it out? [0:54] What ends up happening is you have to reinvent your workflow from scratch, [0:57] and the only way to do that is try it on the problems that you're [1:00] actually working on. [1:01] I think for us, we've been seeing over the last couple of years, [1:06] you watch all the benchmark progress, and when benchmark progress happens, [1:09] you're like, the benchmarks are measuring, is it going to take my job away? [1:13] The higher it gets on a benchmark. [1:15] Our perspective is if you use these tools well, they are good at extending you [1:22] and your power and your influence. [1:24] There needs to be ways to both measure that and to help people understand [1:29] how to do that. [1:30] Because, yes, if you look at benchmarks, they can do pre-AI tasks [1:34] at or better than humans in certain cases. [1:36] But if you're thinking about, how do I use this for my work? [1:39] That opens up a whole set of options. [1:42] We start to see that in engineering first because there's verifiable [1:46] rewards in programming. [1:48] Now you have 10X, 100X engineers who are orchestrating teams of agents to pull [1:53] forward their roadmap by like 2 years. [1:55] I think that is coming to the rest of knowledge work, but it's taken longer, [2:01] A, because it's just harder to know whether or not it's doing a good job, [2:05] and B, knowledge work it's a different way of thinking that maybe developers are a [2:10] little bit more used to, but we think that that's coming. [2:13] We're trying to both discover it and disseminate it, [2:17] so everyone can work that way. [2:18] For people who don't know, I'm curious: what is Every? [2:22] Tell us about what you guys do. [2:23] Every is a media and technology company. [2:25] We build the only subscription you need to stay at the edge of AI. [2:28] For our subscribers, we do a series of education and equipment [2:32] that helps them work at the edge. [2:34] On the education side, we have everything from a newsletter to courses. [2:39] When new models come out, we do vibe checks. [2:41] On the equipment side, now we have an agent that allows you to [2:44] work in the way that we do in public with your company. [2:47] Okay, so you guys built an AI coworker to help solve this problem. [2:51] Tell us about it. What were your goals? [2:53] How's it going? We built the Every Agent. [2:55] It is an AI coworker. [2:57] It's an agent that is the easiest way to AI-pill your whole company. [3:01] We started basically in the OpenClaw craze. [3:05] We started using OpenClaws, and we were like, "Oh, my God, [3:07] this is the coolest thing," and it spread throughout the whole company. [3:09] You had the Mac minis. [3:10] We had all the Mac minis. [3:12] Then, unfortunately, all the OpenClaws died off because [3:14] they're just really hard to maintain. [3:16] We started just using one company agent. [3:18] We were like, "Wow, this is actually really great," and [3:20] noticed how easy it was to spread workflows that we discovered [3:24] throughout the whole company. [3:25] Because our job inside of Every is to work at the edge, but you'd be surprised how [3:29] hard that is to make working at the edge even across the entire company. [3:33] That started to happen. [3:34] We started to see these workflows start to spread. [3:36] What we've also realized is there's a certain type of person that loves Every, [3:39] who we've been calling an AI ambassador, who has that feeling about AI where [3:45] they're like, oh, my God, I have this Codex setup or this Claude [3:48] setup or whatever that is absolutely 10Xing the way that I do my work, [3:52] and I can't imagine working without it. [3:53] But none of my coworkers have any idea what this is, and I'm trying to explain to [3:57] them all the time how they should use it, but they're tuning me out. [4:00] And the beauty of a Slack agent, the Every Agent in particular, [4:03] is that you can show instead of tell how to use this stuff to actually do [4:08] really great work because you're doing it in Slack; you're prompting in [4:11] public instead of in private. [4:13] So we built the Every Agent to have all the things you would expect [4:15] in an agentic coworker. [4:17] It can help you do recurring tasks like gather reports, or do social posts, [4:21] or in any of the kinds of things you might expect. [4:23] But it happens inside of your company's Slack so that your coworkers just [4:27] automatically pick it up, and it starts to spread your best practices. [4:29] We try to empower the ambassadors as much as we can with a tool like this. [4:32] Awesome. [4:33] What's one of your favorite examples of something you've seen somebody do with the [4:38] Every Agent that maybe they could do before, would have taken them longer, [4:41] maybe they couldn't have even really accomplished that thing before? [4:44] I really think that having the Every Agent, having an agent in your [4:49] Slack, there's this crazy top-end stuff. [4:52] But the stuff that I want to focus on is actually there are all these small little [4:57] things that you're doing day-to-day that take up a lot of people's time. [5:00] Having a company agent makes really easy. [5:03] A simple example: we just had open enrollment. [5:05] We did the entire open enrollment process through an agent, through the Every Agent. [5:09] The open enrollment, I'm sure you know, you get a PDF that's like, [5:13] "Here are all these plans," and doing the algebra to figure out the [5:15] plans is actually very hard. [5:17] As part of that process, the Every Agent both knew all the plans [5:21] and knew who you were and helped you figure out which one you should [5:27] select for your particular situation. [5:29] That's something that would have been very hard to do and coordinate [5:32] without a central company agent. [5:33] The Every Agent makes it really easy for A, leaders of the company, [5:36] to create a set of skills that work in that way, and then B, [5:40] get that disseminated to everyone in the org and make sure that the open enrollment [5:44] process actually happens smoothly. Yeah. [5:46] The amount of time you probably saved with not every single person [5:49] pasting separately into Claude. Oh my God! [5:50] Like, "This is the PDF. [5:51] What do I pick?" That saves you a lot of time. [5:53] Honestly, even for me, anytime I get one of those things, [5:57] I just call my dad, and he's like, "I can't believe you're 35. [6:01] I can't believe I have to still explain this to you." Because it is [6:03] actually really complicated. [6:05] Yes, I could just individually paste it into Claude. [6:07] But knowing that this is coming from the company and the company has thought about [6:12] for each person, "Here's what you might want to pick," it really helps [6:15] me feel confident in my choice. [6:16] Yeah, now you can call your dad for more fun reasons. [6:18] Yes, exactly. [6:19] "Sorry, Dad, I automated that task." It's not one that he relishes. [6:24] He didn't like it anymore. Sorry. [6:26] We also have a lot of pretty rote tasks that go along pretty [6:29] with the editorial arm. [6:31] We've spent a lot of time building out something we call Kate Bench, [6:34] which is named after our editor-in-chief, which is mostly like, [6:36] "How do we do top edits before things go out?" There's still a question of [6:41] when to trigger these top edits. [6:43] It's like we flood Kate's inbox. [6:44] It's now much easier to just be to tag the agent in Slack and say, "Hey, [6:48] just review this." The nice part is Kate can see that, and she can judge, "Okay, [6:53] is the agent doing the job or not?" [6:55] This is a really important thing that I think Kate Bench [6:59] gives you a view into how this stuff is going to be used in [7:03] knowledge work going forward. [7:05] Like Willie said, Kate's our editor-in-chief. [7:07] I've been trying to automate her for many years, but in the most [7:11] loving way possible. [7:11] I was going to say, maybe don't have her watch. [7:13] Yeah. [7:13] Automating, I think, sounds scary, but when you actually see it in practice, [7:17] you're like, "Oh, this is actually helpful for her because we publish so much." [7:21] When Kate started, there was four people at Every. [7:25] She was the fourth person, and now there's 30 people. [7:28] She has incredible taste for copy edits, for making sure that every single period, [7:33] every single verb, every single thing is in the right place. [7:36] But we have 10 times the amount of things going out today than we did when she [7:40] started, but she still has to do it all. [7:42] What I did was I just took 30,000 of her historical edits and turned [7:47] it into a skill that is in the Every Agent. [7:51] Now every time we are publishing something, whether it's a piece, [7:54] or a landing page, or an email, anyone in the organization can just say, [7:57] "Do a Kate pass on this," and it will take her taste and apply it, [8:01] literally make edits in the Google Doc with suggested changes as her. [8:06] Then they can accept and reject it. [8:08] What she does is she just goes in and looks at it, and it has most of the things [8:12] that she would do and even catches things she might not. [8:15] She has to spend less time on each piece because it has done a lot of the work. [8:18] Then what we do is we track what is the extra work that she added to this? [8:23] Then we compound it back into the skill so it gets better and better over time. [8:27] She gets to see a graph of, "Here's how much time it's saving you. [8:30] Here's also a rule we might want to add to this so that it catches more things [8:34] that you didn't necessarily know about." [8:36] I think what that does is it flips this automation debate from, [8:40] "Is it going to take away my job?" To, "Actually, I'm an expert [8:44] in an organization. [8:45] My time is taken up all day by having people ask me to do things [8:49] for them that only I can do. [8:51] As the organization scales, in order to scale my impact, previously, [8:54] I would have had to spend more time." [8:57] If you have an agent that compounds like the Every Agent, [9:00] the opportunity for experts inside of organizations is you can spread your [9:03] impact and spread the way that you do work without having to spend more time because [9:07] you're working on a system that does the work for you instead of [9:10] doing all the work yourself. [9:11] Yeah, it's doubly valuable. [9:13] Kate gets more time to do other things, maybe the hardest copy edits, okay, [9:18] and then everybody else gets to take advantage. [9:20] I actually would love to talk about Kate Bench for an example. [9:25] You guys talk a lot about being really bench-pilled and being very on top of, [9:31] "Here's how we do some really cool use case and then measure how well we're [9:36] performing on that really cool use case." [9:38] Tell us more about that mentality and how, as you have people going out there and [9:42] starting to use the Every Agent, they can bring this mentality [9:45] in on continuous improvement. [9:47] Yes, we have really started to get good at this, get good at evals and [9:52] trying to measure these things. [9:53] I think we started with everyone who's in AI and follows us closely knows that [9:58] that doesn't really mean that much. [10:00] The question is not how good is it at a generic benchmark? [10:03] It is how good is it at the work I do? [10:06] That's a totally different way of thinking about whether AI is good [10:11] because it's much more personal. [10:13] That's where we started with our vibe checks, where before a new model comes [10:16] out, we get access to it early, and we actually test it by hand [10:19] across a lot of our workflows. [10:21] Then on the day it launches, we can tell you, "Here's the work we do. [10:24] Here's our taste. [10:26] Here's what we think of this model." [10:27] People rely on that to be like, "Well, I trust Every, and I trust their taste," [10:31] or like, "There's a certain writer at Every that I know has [10:33] a similar taste to me. [10:34] If they like this model, I'll try it." Problem is that that's [10:37] very manual and takes a long time. [10:38] What we started to do is capture those tasks that we do. [10:41] Anytime one of us is doing work with a model day to day, and we notice, "Oh, [10:46] we did something I like," or, "I did something I didn't like," we can [10:49] capture that in one place and then talk to this platform that we built that will [10:54] eventually be part of the Every Agent. [10:56] Talk to this platform that we built about why we liked it and why we didn't. [10:59] That turns it into this set of checks, which are basically unit tests or rules [11:03] for, "Here's the thing you look for." So for Kate Pass or Kate Bench, it would be, [11:08] "Is the headline sentence case?" would be an example. [11:11] What that allows us to do is run many different models on the same scenario and [11:18] then be able to see, for our real work, did this model do a good job or not, [11:21] according to my taste. [11:23] We've started to be able to add this quantified layer to vibe checks that I [11:27] think looks like a benchmark, but actually measures things that you care about. [11:31] Awesome. [11:32] Going back to form factor a bit. [11:34] You talked about, obviously, being in Slack is really powerful. [11:37] All the knowledge workers can have really easy access to this. [11:41] How did you land on Slack? [11:42] How are you thinking about that form factor? [11:44] We actually started with a model of Every, one agent per person available on whatever [11:50] surface you really wanted that agent to be there. [11:53] It was Telegram, it was Slack. [11:56] Part of the reason we, over time, went from a multi-agent setup to really a [12:01] single agent setup is we found that agents got better, being the same dynamic [12:05] the more you invested in them. [12:07] So you had a lot of folks who were like, "Oh, I'm investing in my agent," but those [12:11] gains weren't really spreading out to the rest of the team. [12:14] The places where we really saw agents being able to do real work was when there [12:20] were multiple people investing in them, making better skills, [12:23] giving it better guidance. [12:24] It just made so much more sense for everyone to be investing in a single [12:29] agent rather than everyone having their own agent spread around. [12:33] And Slack was the very natural form factor for where are we when we're [12:37] having these conversations. [12:38] With individual agents and when we're doing collaborative work, [12:41] there's a lot of copy and paste. [12:43] I'm like, "We're going to have a conversation. [12:45] I'm going to copy this over to my Claude Cowork, and I'm going to copy the [12:49] response back, and then you're going to do that." And instead, [12:51] it's much easier to just have that conversation in the same place that your [12:56] coworkers are and the same place your agent is. [12:58] Perfect and easily accessible. [12:59] Let's talk a bit about how you built Every Agent. [13:02] You build this product on Claude Managed Agents, which we're really excited about. [13:06] Tell us, just starting out, what were some of the main reasons why you [13:09] chose to start building on Claude Managed Agents? [13:11] We learned a lot of painful lessons with this first version of [13:13] one agent for everyone. [13:15] One of the main ones is infrastructure is hard. [13:20] A lot of what we want to focus on is the future of work and how to make a better [13:24] product for that future that we see coming. [13:27] We don't necessarily want to spend that time like, "How do we [13:30] orchestrate sandboxes?" [13:32] And so we had this very painful arc where we were like, "Okay, we're going to [13:37] build one agent for everyone." And then it just got unwieldy very quickly because [13:41] you're trying to improve a product and have a strong, durable [13:44] infra layer underneath. [13:46] We took those lessons and turned them into the Every Agent. [13:51] One of our first things was we don't want to be an infra team. [13:54] We want to focus our energy in the place where we think the future is going. [13:58] Claude Managed Agents was great. [14:00] It gave us sandboxes, it gave us memory, it gave us session control, [14:06] all these primitives that really allowed us to focus on the next layer of how do [14:12] we do interaction design for an agent. [14:15] We can just shuffle it off. [14:17] It's like, "Okay, this is being taken care of. [14:19] I'm not getting alerted." [14:21] I am. [14:24] Thank you for your service. [14:28] It's been great. [14:29] Yeah, that's one of those classic things in startups where you're like, "Okay, [14:33] we're going to build an agent platform. [14:35] Should we do the infra layer or not?" And you're like, "Oh, infra [14:37] is not that hard anymore. [14:39] We'll just have Claude spin up a bunch of servers and manage the servers [14:42] for us and blah, blah, blah." [14:44] I was saying that, at least, and Willie was like, [14:45] "It's hard." And I was like, "We have to try it." And then a [14:48] month later, I was like, yeah- [14:50] This is hard. [14:52] Yeah, Claude Managed Agents; it takes something that you don't know is hard, [14:55] but when you do it, you're like, "This is so complicated." It [14:58] makes it very easy for us. Nice. [14:59] That's awesome. There's infra management. [15:01] That's obviously helpful. [15:02] What are some of the other features? You mentioned memory. [15:04] What are some of the things that you've used out of the box with Claude Managed [15:08] Agents that maybe just helped you build agents that were more powerful than [15:13] you would get from just hand rolling a harness and going from there? [15:17] I think the most honest answer is just it allows for the combination of primitives, [15:23] and it allows us more time to do more experiments in the hard area. [15:27] For example, having the ability to isolate out. [15:30] We want this particular managed agent as an edge, like our staging, [15:35] our frontier case, and being able to take the memories and [15:41] the durable state that we have from production [15:45] and copy it over and run our experiments. [15:47] Okay, what happens in this scenario? What happens in this scenario? [15:50] How do we bring the token cost down for our customers? [15:54] There are a million conversational edge cases that you're trying to sort out, [15:57] particularly around stuff that's in the gray area. [16:01] When you're building an AI coworker, there are a lot of things you take [16:04] as an assumption for coworkers. [16:06] They're going to make a good judgment call, and it's hard to [16:10] instruct models how to do that. [16:12] You're running some combination of evals, but also a vibe-based, "Does this feel [16:19] like a real coworker or not?" Experiments. [16:21] Being able to, in terms of a dev experience, be able to slice that off [16:27] and run those experiments in isolation without having to impact our main [16:30] production instance, but also running on the same infra as our [16:33] main production instance, speeds up the team and results in a better product. [16:37] Yeah, that makes a lot of sense. [16:39] And there's a lot going on right now in terms of security with agents, [16:42] and a lot of energy is going into, how do we make sure that [16:46] agents are doing the things we want them to do and not the things that [16:49] we don't want them to do? [16:50] How do Claude Managed Agents help you think about security with Every Agent? [16:54] Yeah, a lot of it is just the built-in isolation. [16:56] It makes it very easy to isolate out the sandbox, the tool calls, [17:03] what's happening, agent in the loop, and what are the environments, [17:07] and where are the memories? [17:08] Because we can control all that, we can make choices around [17:13] who accesses what, when. [17:15] Session management is great for this. [17:17] And that really gives us the nuanced tools to say, "Okay, if Dan requests something, [17:23] does he get the same thing that I would get if I requested it?" And underneath the [17:28] hood, there's a lot of playing with the primitives of what memories [17:32] do we attach to that request? [17:33] Is there other sandboxes the same or different? [17:37] Just having those primitives is just helpful from an engineering standpoint. [17:41] One thing that I do want to add to: I think this relates to security, [17:44] but it also relates to performance, which is there's not a clear separation [17:48] between the model and the harness anymore. [17:51] A model that doesn't come with a computer that it knows how to use [17:56] it's just not as performant. [17:59] Having all of that built in one place by one provider ensures that the model [18:05] knows how to use that stuff really well. [18:07] The day we're filming this, there's a new model launch from you guys, [18:12] and it's one of those things Willie was telling me, if you look at the prompting [18:15] guide, there are some breaking changes that teams would have to care about, [18:20] but in Claude Managed Agents, it's just a 1-line update. [18:23] That's the thing that you get for free that I think is really powerful. [18:27] Yeah, that makes a lot of sense. [18:28] Some of those breaking changes, and just some of the how do we get the [18:32] best out of Claude and make the best use of all the things that it can do? [18:36] As part of releasing the model, we're baking all those things into the [18:40] Claude Managed Agents harness, and so that's cool to hear. [18:42] How did you think about auth and identity? [18:45] Identity for the Every Agent. [18:48] Is it acting on behalf of humans? [18:50] Is it acting on behalf of itself? [18:52] How did you think about that? [18:53] Yeah, there's a lot of nuance to this because as soon as you [18:56] put an agent inside of a company internally, [19:01] people think of it like a human, right? [19:03] And they ask it to do things that fall into these gray zones. [19:06] Like, "Well, a human would know that there's no harm in me being able to access [19:11] this report." But that's hard to encode into a piece of agent logic. [19:17] But we settled on that as our North Star. [19:21] Let's try and make it act as close to a human as it can within Slack. [19:27] The way we do it is when it's in Slack, it acts as a separate entity [19:31] where you're interacting with it. [19:33] Underneath the hood, it has isolated memories and an understanding of [19:38] who you are and what you have access to within the company. [19:42] It's a conversation between two separate things. [19:44] When you ask it to go out and grab something via a tool call or a connection [19:48] or whatnot, it, for the most part, acts on your behalf, [19:51] where it's as an extension of you when it goes into go grab a document [19:57] or look up something. [19:58] And we found that model to be pretty accurate to the way people want to use it, [20:05] where they're fine with this boundary around internal versus external. [20:09] Nice. That makes a lot of sense. [20:11] We touched a little bit on continuous improvement, memory. [20:15] How are you guys thinking about that? [20:17] What is most important to you when it comes to the agent can do certain things [20:21] today really well, maybe not as well as you want? [20:24] How are you thinking about get that agent really great over time at all [20:29] those things better and better? Yeah. [20:32] One of the nice things is the agent is able to see all of the little moments [20:37] where you are adding your taste into request. [20:40] These little nudges of like, "You didn't write this well." [20:43] One today was like, it told a bad joke. [20:45] I was like, "This is not a great joke. [20:46] Give me another joke." We take humor seriously. [20:48] Yeah. [20:52] Those little nudges are really the foundation of what is your taste? [20:55] It's hard to describe your taste all in one big go, but it's easy to take [21:00] it from a thousand examples and distill it out. [21:02] And so those form the foundation of how we build skills, as Dan was [21:07] saying, particular jobs? [21:08] If you're a person who does a lot of editing, the way the agent gets better at [21:13] editing is looking at your 30,000 edits and using that as the foundation for like, [21:20] "Oh, this is the taste I have when it comes to editing." [21:25] And you can think about this for PowerPoint generation. [21:27] You can think about this for email. You can think about this. [21:29] It happens in coding as well, but this is a new frontier for a lot of [21:33] other non-coding knowledge work where you're accumulating these examples. [21:37] The agent is really the repository for all of those. [21:40] I think a good way to think about this is I've been talking a lot about benchmarks, [21:44] but I think it's so important. [21:46] The way that we think about it is starting to shift where think about hiring or [21:50] using an agent as hiring for a job. [21:52] If you're hiring a job candidate, and you're looking at their SAT scores, [21:55] that's a little bit like a benchmark score. [21:58] Directionally, that actually helps a lot. [22:00] So if I'm hiring a job candidate, and they got a 1600 on their SATs versus a [22:03] 300, I would probably take the 1600 person. [22:06] I hope so. Probably. [22:07] But if everyone got a 1600, which is where we are with these models, [22:11] there's a lot more I need to do. [22:13] What I actually want is not your SAT score. [22:15] I want a reference check. I want a work trial. [22:19] For these models, when they come into your company on the first day, [22:22] it's their first day of work. [22:24] They're a new grad who's really smart and has studied a lot, [22:27] but they don't know how you work. [22:29] What we want to try to do is create a system to, as they get experience working [22:33] with you, to compound their knowledge of how you work and what you care about so [22:38] that on day 10, they're better than on day one. [22:41] What we are starting to build is a way to actually prove that. [22:45] To allow you to gather cases where the model did good work or didn't do good [22:50] work, pull out your taste, and then allow it to probably improve [22:55] at those things over time. [22:56] That won't be in the agent on day one, but I think that's where we're going. [22:59] I think generally where work is going, where instead of doing each thing manually [23:04] every single time, you're working on the system. [23:06] Hopefully your agent, whether that's the Every Agent or Claude [23:09] or any agent you're working with, enables you to do that and enables [23:12] you to see how it's getting better. [23:14] Yeah, makes a lot of sense. [23:15] What are your next couple of ideas for how you want to improve the product? [23:18] It's a good question. [23:20] The big one is this compounding system that allows anyone to [23:26] basically create their own little personal benchmark for, "Here's how I work and what [23:30] I care about and what I think is good about this model's response versus another [23:35] model's response," and allows you to see which one you might want to use over time, [23:40] and allows you to build and optimize skills that are for you and based on your [23:45] taste to do your work better over time, and allow that to spread throughout the [23:48] org so that other people can use those skills and work like you can to enable [23:52] more experts to get more work done without spending more time. [23:56] I think that's where we're going. [23:58] We do tons and tons of experiments. [23:59] I have this little experiment called Hands, which is a computer-use agent on [24:04] my desktop that the Every Agent can then go use my computer because you're always [24:09] getting requests where it's like, "Oh, can you do this thing?" I'm just @Every, [24:13] ‘use Hands to do it on my computer.’” [24:15] That's really cool. [24:16] But we're experimenting with a lot of things. [24:18] This is one of the number one questions we get from our customers. [24:23] It's just like, how does this compare to working with Claude Code or [24:26] Claude Cowork an agent? [24:27] We see them as very, they work together really well because of [24:32] experiments like Hands, where there's like, sometimes you want to [24:34] be with coworkers and doing this thing and doing it in public. [24:37] Then sometimes you want to just shift that work very seamlessly over to like, [24:40] "I just want to do it alone," just like you would as if you were human. [24:44] I mostly want to do it alone. [24:45] Really, it's this whole everyone just figuring it out. [24:52] Where do I want to spend my time with this amplified intelligence? [24:58] How do I go back if I don't… [25:01] for us, doing context, how you do context management is both [25:06] gotten more invisible but also more visible as we've moved into… [25:10] personally, it's like, I don't really think about context management anymore. [25:14] Yeah, thank you so much. [25:17] But with coworkers, it really is. [25:20] It's like, "Okay, we're trying to…" Again, if you model it as a coworker, it's like, [25:24] "Oh, there was this conversation over here. [25:26] Were you part of it? [25:27] Do you have enough of an understanding of human [25:31] conversational norms to be like, 'You should interject over here,' or maybe [25:35] you should take these two ideas from separate threads, and they're relevant, [25:39] even though one was a brainstorm and one was something a little [25:41] bit more directional?" [25:42] A lot of the interaction design or personality design, I think, [25:46] is a huge thing that's coming. [25:48] That is such a big thing because Claude, so far, I think this is starting to [25:53] change, but so far, every time it's used to, [25:56] every time it sees a message, it's supposed to respond because you're [26:00] prompting it one-on-one, usually. [26:02] Claude's in a social context with other humans and probably other Claudes. [26:07] There's a whole new set of norms that it doesn't necessarily know out of the box [26:11] that I'm sure you guys will start to improve. [26:14] But we're also thinking about because there's also different styles. [26:17] Do you want it to be an extrovert? [26:19] Do you want it to be an introvert? [26:20] Exactly when and for what use cases would this agent interject versus not are all [26:26] really interesting, complex questions that I think will [26:29] be a lot better in 6 months to a year that we're all going to have [26:33] to work on making happen. [26:34] I love the idea that Every has a certain brand personality, way of working, and [26:39] you're baking that into the Every Agent. [26:42] That personality will come through in the agent. [26:46] One of the design criteria for us was if you're working with the Every Agent, [26:51] it should feel like you have someone who works at Every working with you now. [26:55] That's a really interesting question and a really interesting way to think [26:59] about designing a product. [27:01] As very AGI-pilled people, I'm sure you're thinking a lot about where [27:05] things go from here: where are we going to be in 3 months, 6 months, a year, just in [27:10] terms of how people are going to do work. [27:12] What do you think is going to happen? [27:15] Where do you think that Every Agent is going to need to be? [27:17] What are you going to need to build into it so that it's ready for that moment [27:20] of how work is going to change? It's a very good question. [27:23] I think it goes back to exactly what we're talking about in the beginning, where [27:27] you see developers starting to work in loops, starting to, [27:30] instead of writing every line of code by hand, you're directing sometimes a whole [27:35] organization of Claudes to do your work for you. [27:39] Your job is to make sure that even if you're not doing every unit of work, [27:43] you've created a system that feels like it's an extension of you. [27:47] I've been thinking about words for this. [27:49] I think one word is like fidelity. [27:50] Like the agents have fidelity for what you want, what you care about, [27:54] so that you don't have to be involved in every single thing. [27:57] That's starting to be the case in how we do programming. [28:01] Obviously, with all the security stuff going on, sometimes it's not the case, [28:05] and that's an interesting one. [28:06] It's also something that benchmarks don't really measure that well because, [28:10] by definition, they are more generic. [28:12] They're not about you and what you want. [28:14] But I think over the next year or so, in order to take that next leap in [28:20] knowledge work, we're going to have to make agents that [28:23] have fidelity to what you want and what you care about. [28:26] I think what that will unlock is for individual knowledge workers, [28:31] the same thing that you see in coding, where you're extracting yourself out of [28:35] the loop, you're working on the loop, you're working on the system, [28:39] and making sure the system operates according to what you care about and what [28:44] your taste is and what your vision is. [28:45] You're intervening at certain times, just like a manager would, [28:49] to make sure things were on track. [28:51] What that will do is allow you, A, as an individual person, [28:53] to do way more than you could have before, and then B, allow other people, [28:57] whether it's in your organization or your clients, to use that, use your taste, [29:02] do things like you would without taking up more of your time. [29:05] Yeah, that makes a lot of sense. [29:08] What are you going to need from us, from Anthropic, from the Claude platform [29:12] in order to make Every Agent what it needs to be at that time? [29:16] We're starting to reach the point where successive gains in model intelligence are [29:21] not necessarily able to be consumed, even by power users. [29:25] I think it's so amazing that there are these advances in frontier math or [29:29] now biology, and those are super, super important. [29:32] But I don't want you guys to forget about us knowledge workers, [29:35] because I truly do think even as those gains in those other areas are [29:41] happening, there's still so much low-hanging fruit in [29:44] terms of knowledge work. [29:45] When I get a response back from Claude, do I understand it? [29:48] Does it reference things that are at my level of knowledge or [29:52] that I would remember? [29:53] All those kinds of little things, I think, are so critically important to making a [29:57] tool that people can use every day that I think you will improve [30:01] on those dimensions. [30:02] I expect if you do, there will be a lot more that opens up that allows [30:07] knowledge workers to get more done. Yeah, exactly. [30:09] It's the intelligence, but it's also the speed and the cost [30:12] and all these things that are paired. [30:13] I was going to say the personality. [30:14] We do a lot of thinking and work around personality design. [30:18] Particularly for us, keeping it because we sit between managed agents and our users, [30:23] how do we keep the personality both consistent but also configurable? [30:27] Any tooling there or any little knobs that we could turn to be like, "Okay, [30:32] one customer really wants an aggressive personality, one customer really wants a [30:36] patient personality." It's very standard in the way we can turn Claude into that. [30:42] Yeah, nice. [30:43] I'm excited to go back to this team and talk about new API parameter, [30:46] personality aggression. [30:49] Low, medium. Awesome. [30:51] What's your hot take? [30:52] My hottest take is there's all this discussion about how AI is changing work, [30:58] and there's all this fear of it making work totally go away. [31:05] My experience, or our experience, I think of Every as being a little [31:09] bellwether for how knowledge work might change because everyone is trying to use [31:14] agents as much as possible to get as much of their work done as possible. [31:19] What we found is the more we use agents, the more work there is to do. [31:25] Certainly the work changes. [31:28] Certainly, the way this will play out in different [31:31] parts of the economy in different companies is going to be different. [31:34] But I think that, that is a general truth that when you reach this horizon of work, [31:39] you expect there to be nothing, but actually it just opens up a [31:42] whole new horizon in front of you. [31:45] That provides an opportunity for everyone to actually just do better work at a [31:50] higher quality that feels like them. [31:52] Automating your work doesn't mean that it's done in a robotic way or that [31:57] you don't pay attention to it at all. [31:58] It just means that where you're working in the system has changed. [32:02] Okay, last fun question. [32:03] I think a lot of people have this magic fun moment that they sometimes feel [32:08] with agents that can be really powerful. [32:10] One of my favorite ones is the emoji that I love to use the most is [32:14] the blue-colored heart. [32:15] One time I was working, and I was working with Claude, [32:18] and Claude showed up and reacted to one of my messages with a blue heart. [32:22] I was like, "Whoa, how did Claude know?" I'm so curious for you [32:26] guys with Every Agent. [32:28] What's been just the most fun magic moment that you've experienced [32:31] in working with it? [32:32] Our CEO keeps track of basically beers he owes folks. [32:38] Like, "Oh, my bad. [32:40] My bad. I did this. [32:41] Every, remind me that I owe Dan a beer for this." The agent will insert that back in [32:49] occasionally in the appropriate conversation where it's [32:52] just like, "Oh, yeah. Remember when Dan hooked you up? [32:56] Have you reclaimed that beer? [32:57] Have you claimed that beer yet?" [32:59] The lead on the Every website, his name is Andrey, [33:03] and he does these really funny announcements that are just, [33:07] they're in all caps, and they're like, "You must turn on two-factor [33:11] authentication in all of your logins," or whatever. [33:13] But he spends a lot of time making them just really funny. [33:16] I just made a skill that was like, "Have the Every Agent do announcements in [33:20] his voice." And so now people are just using his announcement skill to [33:23] do all of their announcements. [33:24] I was like, "It's just very pleasing that that just works." [33:27] Poor Andrey. How will people take him seriously? [33:31] Nice. That's awesome. [33:32] Okay, well, I'm excited for the day that the Every Agent orders you guys [33:35] some beers to pay the debt. This was awesome. [33:38] Thank you, guys, so much for spending the time. [33:40] Thank you. Thank you so much.