WTF Is the Context Layer? The Missing Infrastructure for Production Agents — Prukalpa Sankar
频道: AI Engineer
视频: https://www.youtube.com/watch?v=8G_1-3IO4ZQ
原文语言: en
统计: 共 9 轮 · Prukalpa 9
[0:12] Prukalpa
Hi everyone. Uh my name is Praalpa. I'm the founder of Atlan. Um and uh today I'm going to talk about this thing where context is having its moment. Uh and so my goal today is to talk about like WTF is the context layer. Um, just before I start, and I promise this is the last time. Um, I don't know if the clicker is working. Atlin, we it's it's working. Yeah. Thank you. Um the problem we solve is we say AI doesn't know your business. We fix that. We work with an incredible group of companies around the world ranging from GitLab and Zoom and Discord and Affirm to large enterprises like Mastercard and General Motors. Um and about a year ago, uh my co-founder and I went on stage and we said, uh at the dawn of the internet era, Bill Gates had written this very famous blog post and it said content is king. Um and as we or the dawn of the agentic era, context will be king. Um since then it feels like 2026 is the year of context. context graphs anyone um uh you know every every two days you see some version of context uh popping up and so what is going on um I believe the answer to this kind of is in this reality distortion field that we live in uh I live here in the Bay Area every day or two I have conversations with people which kind of go like how far are we from AGI and we have a debate and we're like well one year three years so on uh There is no doubt that the models are getting exponentially smarter by the day. Uh two years ago they couldn't pass the bar. Today if they were to take the bar it was they're the top 1% of test scorers. On the other hand they're not exponentially more useful by any benchmark. Uh one out of five you know AI use cases actually make it to production. U you know 56% of CEOs say that there's zero financial benefit from AI today. So what's going on? I believe hidden in plain sight is actually um how performance is measured in the human world. Uh cognitive intelligence doesn't really determine real world effectiveness. Uh in fact only 10% of job performance variance is explained by IQ. Like just think about it.
大家好，我叫 Prukalpa，是 Atlan 的创始人。今天我想聊的话题是：context（上下文）正迎来它的高光时刻。所以我今天的目标就是讲清楚——上下文层（context layer）到底是个什么东西。开始之前先说一句，我保证这是最后一次问：这个翻页器能用吗？好，能用了，谢谢。我们解决的问题一句话概括就是：AI 不了解你的业务，我们来解决这件事。我们服务着一批非常出色的公司，从 GitLab、Zoom、Discord、Affirm，到 Mastercard、General Motors 这样的大型企业。大约一年前，我和联合创始人上台讲过一句话：互联网时代拉开帷幕时，Bill Gates 写过那篇著名的博客，说 content is king（内容为王）；而在 agentic 时代的黎明，context 将为王。从那以后，感觉 2026 年成了 context 之年——context graph 什么的，隔三差五就能看到各种 context 概念冒出来。那到底发生了什么？我觉得答案就藏在我们身处的这个「现实扭曲力场」里。我住在湾区，每隔一两天就会跟人聊到「我们离 AGI 还有多远」，然后争论是一年还是三年。毫无疑问，模型每天都在指数级变聪明——两年前它们连律师资格考试都过不了，今天去考就是前 1% 的高分选手。但另一方面，从任何指标看，它们并没有指数级变得更有用。五个 AI 用例里只有一个真正上了生产环境；56% 的 CEO 说 AI 至今没带来任何财务收益。那问题出在哪？我认为答案其实一直明摆着，就藏在人类世界衡量绩效的方式里：认知智力并不真正决定现实世界的效能。事实上，工作绩效的差异只有 10% 能用 IQ 解释。你仔细想想。
[3:03] Prukalpa
Would you say your smartest um you know teammate who scored the highest on the SATs is also your best teammate or would you say no it's the person who works the most and takes the most feedback and learns the fastest in the real world we care about performance and performance is outcomes that you deliver in the real world and performance is a function of two things it's a function of intelligence which is cognitive horsepower that's what the model benchmarks measure every day But it's also a function of context. This is what they say in the human world as learning on the job, right? Knowledge and skills and expertise that you learn over time. And in the last decade, uh we have compounded on one of those parameters. Uh intelligence has thousandxed in the last decade. Just in the last 6 months, we have 2xed on that axis. On the other hand, context, the situated knowledge of your business, that's barely moved. We've moved some data to the cloud uh but that's about it. It's otherwise logged in dashboards and Slack threads and uh the head of that analyst who might be leaving next week. Um and so the question ahead of us and I really believe this is the next frontier is how do we help AI build context about our business? Um, and every time I'm faced with a question about how do we help AI do this, I always like to go back and understand how did we help humans do this? Uh, so I'm going to take you into the life of, you know, a u exemplar employee Maya. Uh, let's say she's a data analyst at Mech Context Burgers because I thought I was going to be creative and I'm not very creative. Um, and you know, let's say she's that analyst that everybody, you know, pings in your company. Uh, right? She's the person that everybody sends a message to every morning when they're trying to solve a problem. So, let's say this morning, uh, there's a franchisee owner who sends her a message and says, "Why is my drive-thru time up this week? Why is this metric up this week?" Sounds like a really simple question. Um, but it's actually a really complicated question to ask. Just to answer this one very simple question, Maya first needs to know uh what is drive-through time uh and who's asking?
你会说 SAT 分数最高、最聪明的那个同事就是你最好的队友吗？还是说，最好的队友其实是那个干活最多、最能接受反馈、在真实世界里学得最快的人？在现实世界里我们看重的是 performance（绩效），也就是你在真实世界交付的成果。而绩效是两个东西的函数：一个是智力，也就是认知马力，这正是模型 benchmark 每天在测的东西；另一个是 context，人类世界里管这个叫「在岗学习」——你随时间积累起来的知识、技能和专业经验。过去十年，我们只在其中一个参数上疯狂复利：智力翻了上千倍，光是过去 6 个月就又翻了一倍。而另一边，context——关于你业务的情境化知识——几乎原地踏步。我们把一些数据搬上了云，仅此而已。其余的都锁在 dashboard 里、Slack 讨论串里，以及那个下周可能就要离职的分析师的脑子里。所以摆在我们面前的问题——我真心认为这是下一个前沿——是：怎么帮 AI 建立起对我们业务的 context？每当我面对「怎么帮 AI 做到这件事」的问题时，我总喜欢回头看：我们当初是怎么帮人类做到的？所以我带大家走进一位模范员工 Maya 的日常。假设她是 Mech Context Burgers 的数据分析师——我本来想起个有创意的名字，但我实在没什么创意。她就是你们公司里人人都会去 ping 的那种分析师，每天早上大家遇到问题都会给她发消息。比如今天早上，一位加盟店老板给她发消息问：「为什么我这周的 drive-thru 时间变长了？为什么这个指标这周上升了？」听起来是个特别简单的问题，但其实非常复杂。光是回答这一个简单的问题，Maya 首先得知道：drive-through time 到底是什么定义？以及，是谁在问？
[5:15] Prukalpa
Is it finance or is it you know my ops team? And it might mean different things. Uh but not just that, what does this week mean? Is the cutoff period Monday to Sunday? Is it Pacific time? Is it Eastern time? Uh that's knowledge. Like that's facts. That's the map of the business. Um but not just that. Uh there's expertise uh right there's um you know a diagnostic playbook. What what does a great analyst do? They know that you know quarter 3 is a season seasonal quarter because of weather patterns and they know to go check if the reason there's a spike is because of seasonality. They also know that the company launched a product uh just that previous quarter and so they know to check if that's why the root cause analysis failed. Uh this is expertise and skills that people pick up over time as they learn on the job. Uh and then there's norms, right? Um there's, you know, persona scoping. Who's asking the question? How do I answer this question? Um and Maya, she's one of those like cool people. She nails it. She sends an answer not just with the answer, but with the why and the root cause, and she finds the reason for it. How did Maya learn to do this? Um she just joined the company a year ago. Um first Maya you know has for like she joined and she got some training like all of us do but that's not where any of us learn right in our companies. How do we learn? We learn because you shadow like the best teammate and then you see why they're doing something and then you learn from that and then you make a mistake. Who here has learned more from a mistake than anything else? Right? You make a mistake and then you learn. uh you your manager gives you feedback and you learn not to do that again. You deal with an edge case and then you learn from that. That's how all of us humans learn at work. And so then the question is how do you help build the agent Maya? Uh and now I want to walk you through our experiments and learnings as we've built this at Atlan um era one and this was roughly about 18 months ago now. Um we uh started on the the track of bootstrapping agents. Um and the way we went about it was and we started this with our customer experience team.
是财务在问，还是运营团队在问？对不同的人它可能意味着不同的东西。不止如此，「这周」是什么意思？统计周期是周一到周日吗？按太平洋时间还是东部时间？这些是知识，是事实，是业务的地图。但还不止这些，还有专业经验（expertise）——一套诊断 playbook。优秀的分析师会怎么做？她知道第三季度因为天气因素是季节性波动的季度，所以会先去查这次的飙升是不是季节性原因；她还知道公司上个季度刚发布了一款新产品，所以会去查是不是这个导致了根因分析出问题。这些就是人们在工作中随时间习得的专业技能。然后还有规范（norms）——比如按角色定回答口径：是谁在问？我该怎么回答？Maya 是那种特别厉害的人，她答得漂亮：不光给出答案，还附上原因和根因分析，把问题的来龙去脉都找出来了。那 Maya 是怎么学会这些的？她一年前才入职。刚进公司时她跟我们所有人一样接受了一些培训，但我们谁也不是靠培训学会干活的，对吧？我们是怎么学的？你跟着最强的同事「影子学习」，观察他为什么这么做，从中学习；然后你犯个错——在座有谁不是从错误里学到的东西最多？——你犯了错，经理给你反馈，你学会下次不再犯；你处理了一个边缘案例，又从中学到东西。这就是我们人类在职场上学习的方式。那么问题来了：怎么构建 agent 版的 Maya？接下来我想跟大家分享我们在 Atlan 一路构建的实验和心得。第一个阶段（era one），大概是 18 个月前，我们走的是「自举 agent」（bootstrapping agents）这条路，从我们的客户体验团队开始。
[7:38] Prukalpa
Uh and we did this jobs to be done analysis map, right? And so we said, hey, if you are someone on our customer experience team, what are all the things that you do on a day-to-day basis? And then we made some hypothesis. We we said, you know, for example, one part of the job is documentation and meeting prep. Uh we said well AI could probably do that job pretty well. Uh and so we build a scaling factor. So on the other hand relationship management is something that our customer experience team does and we said hm that doesn't sound like something AI is going to be able to do anytime soon. And so we built a scaling factor and then we basically started bootstrapping these individual agents that were like built for that specific topic. Uh our team got creative. So we had Hermione who is our health intelligence lead and then we had you know money penny who was our financial risk analyst and we just made that particular agent really good at doing that one thing. Um and that worked for some time um but then we realized there were some challenges with this approach. The first context engineering uh we got to the point by middle of last year where building an agent was really easy took like 5 minutes. uh but giving it the business context that it took to actually get it to be accurate took forever. Uh quality of the agent often dependent uh on the quality of context engineering and that led to a lot of weird lost trust cases with our stakeholders. Um then as we started taking this into production we started seeing that these agents basically were kind of like living on their own island. Uh now imagine for example if you're in a human team and your marketing changes positioning on your you know and then they come to the town hall and they tell you that they changed positioning and so then you know the SDR on your team or your sales development rep they know that they should use that new positioning. This is like the infrastructure that we've built for humans inside our organizations. Agents didn't have them. So our marketing team had these agents and they started making changes to that and then our SDR agent on our website was still pitching the old version.
我们做了一个 jobs-to-be-done 分析地图：如果你是客户体验团队的一员，你每天要做的所有事情是什么？然后我们做了一些假设。比如我们说，写文档和会议准备是工作的一部分，AI 应该能干得不错，于是给它定了一个放大系数（scaling factor）；另一方面，客户关系管理也是客户体验团队的活儿，我们觉得这事 AI 短期内干不了，也相应做了标记。然后我们就开始为一个个具体主题自举出专属 agent。团队起名很有创意：有负责客户健康度情报的 Hermione，有做财务风险分析的 Moneypenny——我们把每个 agent 都调教得特别擅长干那一件事。这个方法管用了一段时间，但后来我们发现有几个问题。第一是 context engineering（上下文工程）：到去年年中，搭一个 agent 已经非常容易了，5 分钟搞定；但要给它灌进足够的业务 context、让它真正准确，却要花掉大把时间。agent 的质量往往取决于 context engineering 的质量，这导致出现了很多奇怪的失误，让业务方对它失去信任。第二，当我们把这些 agent 推向生产环境后，发现它们基本上都活在自己的孤岛上。想象一下人类团队里的场景：市场部改了产品定位，然后他们在全员大会上宣布这次调整，于是你团队里的 SDR（销售开发代表）就知道要用新的话术了——这是我们在组织内为人类搭好的基础设施。但 agent 没有这些。我们市场团队的 agent 更新了定位，而官网上的 SDR agent 还在推销旧版本的说辞。
[9:50] Prukalpa
Uh we had no idea how any of these things were even connected. So we didn't even know how to like run this as as a team of agents. Uh when an agent gets something wrong, this is hard. Uh it was really hard to like trace back what happened. Was it the model? Was it the agent? Was it the context? Like where how do we even go back and fix this? Um and over time we started dealing with uh context sprawl. Uh we had the the the hard part about this was agents all had their own memory systems to a certain extent. So they were learning they were all learning separately and they were learning differently. Uh it became very very difficult very quickly to say okay what does the single version of truth here look like? Um and then over time we actually went through in the last 12 months we've gone through cycles of at the agentic layer about 12 months ago we were using one of these no code type builders uh called relevance we went from there into Google ADK then we tried glean uh start of this year we moved to cloud code now we are kind of like 50/50 claude and codeex um and every single time as these changes happened uh our context got trapped in each of these individuals systems. Um, so started this year as general purpose agents started to become a thing, we said, what if there was a different approach with general purpose agents. Um, again going back to the human world, well Maya, she's not an individual star. She's part of a team, right? And you know, you talk about these dream teams like Maya and someone who runs customer support and someone who launches ads. These people work really well together. And often these dream teams are built on shared context, right? Uh they have a shared language. Uh they have a shared picture of what's true today. They have shared playbooks. Uh they have shared norms, who's allowed to make what decision. Uh and then they learn together. I think this is the most important part of it. They have compounding learning loops of what good looks like.
我们完全不知道这些东西之间是怎么关联的，甚至不知道该怎么把它们当成一个「agent 团队」来运营。而且当某个 agent 出错时，特别难办——很难回溯到底发生了什么：是模型的问题？agent 的问题？还是 context 的问题？我们要从哪儿下手去修？随着时间推移，我们还开始面对 context sprawl（上下文蔓延）。最麻烦的是，这些 agent 某种程度上都有自己独立的记忆系统，它们在各自学习，而且学的方向还不一样。很快你就说不清「唯一可信的事实版本」到底长什么样了。另外，过去 12 个月我们在 agent 层换了好几轮工具：大约 12 个月前我们用的是一个叫 Relevance 的 no-code 构建器，然后换到 Google ADK，又试了 Glean，今年年初迁到 Claude Code，现在基本是 Claude 和 Codex 五五开。每一次切换，我们的 context 都被困在上一个系统里带不走。于是今年年初，随着通用 agent（general purpose agents）开始成气候，我们想：能不能换一种思路？还是回到人类世界：Maya 不是单打独斗的明星，她是团队的一员，对吧？你会看到那种「梦之队」——Maya 加上负责客户支持的人、负责投广告的人，这些人配合得天衣无缝。而这些梦之队往往建立在共享 context 之上：他们有共同的语言，有对「当下什么是真的」的共同认知，有共享的 playbook，有共享的规范——谁有权做什么决定。最重要的是，他们一起学习，有关于「什么是好」的复利式学习循环。
[12:02] Prukalpa
uh and they have shared memory that you know oh we launched this thing last quarter and it like was terrible and we're not going to make that mistake again right and so we said is there a way to bring that into the way we think about AI in our companies and so the mental model we started working on was we said okay we have these teams of humans and they're across the board and can these people essentially start building domain skills so each of them is responsible for a certain set skills. All of this goes into this common one place which is this one company brain of sorts, right? I like to think of this as the context layer. Uh and then this has a bunch of retrieval mechanisms which then talks to the general purpose agent across the ecosystem. So then we started an experiment. Uh this is some version of what our marketing team ended up building. So you'll see on the left those are all the systems that our marketing team uses. So data systems, our social and community platforms, our ad platforms, our analytics platforms. Um and then you'll see this agent block. Uh we built this very specifically for um having openness. So we had claw code and co-work. We also had our own claw that we deployed which has you know essentially talks in our slack channels. Um and then we used some external products like qualified and artisan. Uh in the middle is kind of this context layer that our team started building. So think of it as our best SEO person was building their SEO skill. Uh our best competitive intel person was building the best competitive intel skill and that kind of became this common repo that we were building into and pulling out from. This sort of became our living brain. Over time, we realized there were some things that we needed in this brain, right? Uh we realized we needed a data graph like if for example our autonomous ads agent, we realized it needs to do analysis on a daily basis. So like which table should I go pull from? Uh we needed a library of skills. We also needed some other things, semantics, metrics, what is ARR, how do you measure that? Uh what is a qualified lead in our company? uh and or structure entities things like that.
他们还有共享记忆——「我们上个季度发布过这个东西，结果糟透了，这个错不能再犯」，对吧？于是我们就问：能不能把这一套带进公司里使用 AI 的方式？我们开始打磨的心智模型是这样的：我们有一支支人类团队，分布在各个职能上，能不能让他们各自开始构建自己领域的 skill——每个人负责一组特定技能？所有这些都汇入一个共同的地方，也就是某种「公司大脑」（company brain）。我喜欢把它叫做上下文层（context layer）。然后这个层配有一系列检索机制，去对接整个生态里的通用 agent。于是我们启动了一个实验。这是我们市场团队最终搭出来的某个版本：你会看到左边是市场团队用的所有系统——数据系统、社交和社区平台、广告平台、分析平台。然后是 agent 模块。我们特意把它设计得很开放：有 Claude Code 和 Cowork，也有我们自己部署的 Claude，它直接在我们的 Slack 频道里对话；还用了一些外部产品，比如 Qualified 和 Artisan。中间就是我们团队开始搭建的 context layer。可以这样理解：我们最强的 SEO 专家在构建 SEO skill，最强的竞争情报专家在构建最好的竞争情报 skill，这就变成了一个大家共同写入、共同取用的公共仓库。它渐渐成了我们的「活体大脑」。随着时间推移，我们发现这个大脑里还需要一些东西：比如需要一张 data graph（数据图谱）——我们的自主投放广告 agent 每天都要做分析，它得知道该去拉哪张表；需要一个 skill 库；还需要一些别的东西，比如语义层和指标定义——什么是 ARR？怎么计算？在我们公司里什么才算一个合格的 qualified lead？还有结构化实体等等。
[14:16] Prukalpa
Over the last 6 months, we ended up creating about 300 skills and 40 agents in this team. Uh which has been incredible. Uh but then with this approach too, we realized that there were some challenges. We realized that context kind of needs to be managed like code. Um so some challenges, let's pick skills. uh dependency management became really complicated. So for example, we have this comparative intelligence skill and it learns from the market on what's changing in the market and it improves. Um it feeds our category positioning skill which then feeds our sales battle card skill. Uh now each of these skills is learning and evolving. Uh but every time they learn and evolve it breaks something downstream. uh and these skills very quickly start getting outdated and start drifting. Uh who owns skill quality became another thing like who eventually owns the quality of this security and governance was a nightmare. Uh we had secrets hardcoded in ENV files. Uh it was people were downloading these public skill repos. This the whole thing was like a nightmare. Um and then I talked about context portability across all these multi-agent systems. I started this talk by saying WTF is a context layer. Uh these are the problems that a context layer is meant to solve. Um the question I like to ask is what does the GitHub for context look like? Um few thoughts. Uh company context needs life cycle management, collaboration and versioning. Uh just like code does. uh you know there's questions like what's local context what's global context how do I keep this updated so on uh some thoughts in this can skills have a profile just like code does uh can that have a self-learning learning loop that's baked into it uh what does quality management look like can you have security and postures posture management associated with that that's really like the first step uh I see this as like having something that has built-in versioning and quality and dependency management. So you should be able to say, "Hey, this thing impacts all these other things. This is the approver. This is the maintainer. These are the contributors.
过去 6 个月，这个团队一共创建了大约 300 个 skill 和 40 个 agent，效果非常惊人。但这套方法同样暴露出一些挑战。我们意识到：context 需要像代码一样被管理。举几个问题，就拿 skill 来说：依赖管理变得非常复杂。比如我们有一个竞争情报 skill，它从市场上学习动向、不断进化；它喂给品类定位 skill，后者又喂给销售 battle card skill。这些 skill 各自都在学习和演化，但每次一进化就会弄坏下游的东西。而且这些 skill 很快就会过时、开始漂移。谁来对 skill 的质量负责？这又成了一个问题——这东西的质量最终归谁管？安全与治理更是一场噩梦：密钥被硬编码在 ENV 文件里，有人在随便下载公开的 skill 仓库，整个局面一团糟。再加上我前面说的、context 在多 agent 系统之间的可移植性问题。我开场问的是「context layer 到底是什么」——它要解决的正是这些问题。我喜欢问的一个问题是：context 界的 GitHub 应该长什么样？几点想法：公司的 context 需要生命周期管理、协作和版本管理，就像代码一样。还有一些问题，比如什么是局部 context、什么是全局 context、怎么保持更新等等。再往下想：skill 能不能像代码一样有自己的档案（profile）？能不能内置一个自学习循环？质量管理长什么样？能不能给它配上安全态势管理？这其实只是第一步。我设想的是一个内置版本管理、质量管理和依赖管理的东西：你应该能够说，「这个东西会影响下游这些东西；这是审批人，这是维护者，这些是贡献者。」
[16:38] Prukalpa
How do you build like kind of human plus AI workspaces that these that these skills uh are managed via? Second thing, every AI interaction creates more context and harnessing this uh is gold. uh uh there's been I know a lot of talks about self-improving loops uh we have found that with traces deploying a specific harness that actually is specialized in being able to go and reverse construct from that. So think of it as AI that's reading through all your traces and almost brings it back to your maintainer loop and says approve reject approve reject improve this over time. Uh that's the compounding learning loop. And the third often a lot of people ask me this question which is like how do I start because my business is like really disperate and I have all these like 60 systems and how do I even start? One of the biggest learnings we've had is context is hidden in these in business systems. Uh and across this context quality can really compound. So for example, if you're able to connect your Salesforce and your HubSpot to your data warehouse to your application layer and then you're able to reverse construct how these things are actually connected one to another, context today gets lost in every one of those hops. But if you can reverse construct that and then deploy AI on top of it, we've seen incredible accuracy in being able to reverse construct the first version of your company brain. So I'll end with this. Uh the way I think about a context layer is it's a system that turns knowledge and expertise and norms that we talked about that Maya knows into a machine usable context for AI systems. Um at a very high level the way I like to think of it is it looks like this. Uh it continually is mining context from your business systems. It's feeding this in to that one company brain. It's harnessing this in skills and context development life cycles as your teams go and deploy these agents. And then it has a bunch of ways you can retrieve it. So MCP, SQL, vector retrieval, hybrid assembly, all these different ways that you retrieve it and pull back from traces and build this compounding learning loop. Today we're largely building agents by hard- coding context.
你要怎么搭建那种「人类 + AI」协同的工作空间，让这些 skill 通过它来管理？第二点：每一次 AI 交互都会产生新的 context，能把这些收集利用起来就是金矿。我知道已经有很多关于自我改进循环（self-improving loops）的分享。我们的发现是：针对 traces（交互轨迹）部署一个专门的 harness，让它擅长从轨迹里反向重建。可以理解为一个 AI 在通读你所有的 traces，然后把结果送回维护者的审核循环，让人「批准、拒绝、批准、拒绝、改进」，如此往复。这就是复利式学习循环。第三点，很多人问我：我的业务特别分散，有六十来个系统，我到底该从哪儿开始？我们最大的心得之一是：context 就藏在这些业务系统里，而且跨系统的 context 质量是可以复利叠加的。举个例子，如果你能把 Salesforce、HubSpot 连到你的数据仓库，再连到应用层，然后反向重建出这些东西彼此之间的关联——今天的 context 在每一次跳转中都会丢失，但如果你能把关联重建出来，再在上面部署 AI，我们看到它能以惊人的准确率反向构建出你「公司大脑」的第一个版本。最后总结一下。我理解的 context layer，是一个把知识、专业经验和规范——就是 Maya 脑子里的那些东西——转化成 AI 系统可用的机器可读 context 的系统。从高层视角看，它大概是这样运转的：持续地从你的业务系统里挖掘 context，汇入那个统一的公司大脑；在你的团队部署 agent 的过程中，通过 skill 和 context 开发生命周期把它管起来；然后提供一系列检索方式——MCP、SQL、向量检索、混合组装等等——同时从 traces 里回收信号，构建复利式学习循环。而今天，我们构建 agent 的方式基本还是在硬编码 context。
[19:07] Prukalpa
The scale of this problem I truly believe is unhived because with scale this can become really unsustainable. Uh and a little dangerous like all of us know this this old joke which is if you ask sales and finance the revenue number you're going to get two different numbers. uh we're fast approaching a moment of starting to deploy autonomous systems where the same thing is starting to happen. So I'll end with one last thing. I started this presentation by saying context is king. Um I'd like to end it by saying context is also IP. Something I think a lot about is in a world where you and your competitor have access to the same models and the same intelligence, what differentiates a company? What differentiates a customer support agent at American Express versus Amazon? Uh that's how you do business. That's what makes your company special. uh context is how we take and encode our culture and our norms into something that we will be proud of as we build autonomous frontier firms. Um and that's all I had. Um you can find me at proalpa on Twitter um or write to me. We are actively working with folks on the frontier ongoing and shipping and building company brands. Um so if you'd like to talk to us, feel free to reach out. Thank you.
我真心认为，这个问题的严重程度被大大低估了。因为一旦规模上去，事情会变得非常难以为继，甚至有点危险。大家都听过那个老笑话：你去问销售和财务要营收数字，得到的会是两个不同的数。而我们正快速逼近这样一个时刻——开始部署自主运行的系统时，同样的事情也在上演。最后再说一点作为收尾。我在开场时说过 context is king（上下文为王），结尾我想说：context 也是 IP（知识产权）。我经常思考一个问题：当你和你的竞争对手用的是同样的模型、同样的智能时，公司之间靠什么来区分？American Express 的客服 agent 和 Amazon 的客服 agent，差别在哪？答案是你做生意的方式——那才是让你的公司与众不同的东西。context 就是我们把自己的文化和规范编码进系统的方式，让我们在打造自主运行的前沿企业时，能为之自豪。我要讲的就是这些。你可以在 Twitter 上找到我，账号是 prukalpa，也欢迎给我写邮件。我们正在和前沿团队持续合作，一起构建和交付。如果你想聊聊，随时联系我们。谢谢大家。