software engineers are going to have to tackle a new set of problems to unlock powerful multiagent systems to solve the world's hardest problems - right now is one of the most interesting moments in history to be an engineer.
软件工程师将要面对一组新的问题,才能释放出强大的多智能体系统去解决世界上最难的问题——此刻,是历史上做一名工程师最有意思的时刻之一。
The new set of problems for engineers to solve
5 min read
A lot of software engineers are going through what can only be described as an identity crisis. The job has changed a lot. Some are depressed, others are having the time of their lives. There's a really specific thing I see happening that makes me really excited for what engineering looks like over the next, call it a year or two - AI just isn't going to be able to create the incredible outcomes it's capable of creating unless software engineers solve problems to make it possible. This is existential, which is honestly a lot of fun, so it's a pretty incredible time to be a software engineer.
工程师要解决的一组新问题
约 5 分钟阅读
许多软件工程师正在经历一场只能被称作"身份危机"的状态。工作本身变了很多:有人感到沮丧,有人则觉得这是生涯中最好的时光。我看到一件非常具体的事情正在发生,让我对未来一两年工程师这份职业抱有强烈期待——除非软件工程师动手把那些问题解决掉,AI 根本没办法真正交付出它能力范围内的那些惊人成果。这是一件涉及"存在意义"的事,坦白说也挺有意思的,所以这是当软件工程师非常不错的时代。
Agents started out as basic, human in the loop chats. Then models got better at running for longer on harder problems, and agentic systems started moving to the cloud. While that was happening, models have started to get better at working together, and people have built basic multiagent systems. Like an orchestrator that spawns some subagents for a given task, or giving agents jobs other than executing to improve outcomes, like grading or advising or dreaming.
Now, some folks are starting to build truly amazing multiagent systems to tackle the world's hardest and most important problems, like serious research, discovery, and hardware design. Picture thousands of agents, persistently working towards these outcomes. This comes with a new set of software engineering problems that we're already seeing crop up.
Infrastructure needs to evolve to handle burstier workloads. We need proper authentication and authorization to identify agents, let them work together, enable them to take the actions they should, and prevent them from taking the actions they shouldn't. Security and proper isolation are critical for untrusted code and content. And we've gotta have monitoring and the ability to stop a system in its tracks when it's veering into the wrong territory.
Models will keep getting better, and these types of multiagent systems will become really powerful. But software engineers have to solve all these problems if we're going to actually unlock that power and enable the world to trust it to create good outcomes like cured diseases rather than bad outcomes.
智能体最早的形态是最基础的"人在回路"对话。后来模型越来越擅长在更难的问题上持续更久地工作,agentic 系统也开始搬到云上。与此同时,模型之间互相配合的能力也在变强,人们开始搭建基础的多智能体系统。比如一个编排器(orchestrator)为某个任务派生出若干子智能体,或者让智能体承担执行之外的角色来改善结果,比如打分、给建议、做"白日梦"。
如今,一些人开始搭建真正惊人的多智能体系统,去挑战世界上最难、最重要的问题,比如严肃的科研、发现与硬件设计。想象一下成千上万个智能体,持续地朝着这些目标工作。这会带来一组全新的软件工程问题,而且我们已经开始看到它们冒出来了。
基础设施需要进化,来承载更"突发性"的工作负载。我们需要像样的身份认证与授权机制,用来识别智能体、让它们协同工作、允许它们做该做的事,并阻止它们做不该做的事。对不受信任的代码和内容而言,安全与适当的隔离至关重要。还必须有监控手段,以及在系统偏离正轨时能立刻把它拦下来的能力。
模型会继续变好,这类多智能体系统会变得非常强大。但只有软件工程师把这些问题都解决掉,我们才能真正把这份能力释放出来,让世界敢于信任它去产生治愈疾病这样的好结果,而不是坏结果。
Infrastructure needs to evolve
The software industry has been building on Kubernetes, which was designed for stateless, replicated apps or stable numbered sets of stateful pods. Agents are totally different. They're essentially isolated, stateful, single-instance workloads. They can go idle for arbitrary amounts of time between bits of work, and they spend a lot of time waiting on tools, model inference, other agents, or human input. Some workloads need fast startup times, others need to avoid wasting compute, and others need crazy scale. Amazing engineers are pushing hard on this problem space - like AI infra companies building new filesystems and sandboxing offerings, and the Google team building Agent Substrate as a new runtime. There's a ton left to figure out.
基础设施需要进化
软件行业一直建立在 Kubernetes 之上,而 Kubernetes 当初是为无状态的可复制应用,或者有稳定编号的有状态 pod 集合设计的。智能体完全不同。它们本质上是隔离的、有状态的、单实例的工作负载。它们可能在工作片段之间任意时长地空闲,又会花大量时间等待工具、等待模型推理、等待其他智能体,或者等待人的输入。有的工作负载需要极快的启动时间,有的需要避免浪费算力,还有的需要疯狂的规模。目前已经有一批了不起的工程师在这个问题空间里发力——比如 AI 基础设施公司在做新的文件系统和沙箱产品,Google 的团队在把 Agent Substrate 做成一个新的 runtime。但剩下要搞清楚的东西还有一大堆。
Identity and access
There are problems to solve in agent authentication - identifying an agent that took an action. This area has the benefit of the building blocks that have been built and hardened from the microservices world, like workload identity with short-lived certs rather than static keys, and folks are beginning to apply these concepts to agents.
Agent authorization as a problem space is much more wide open - what do you allow an agent to do? Many agents work on behalf of a human, and use that human's identity and credentials to access various systems and take actions. There's already a problem in there, which is identifying that it was the agent and not the actual human who took an action. But there are deeper problems in this on-behalf-of (OBO) setup, like consent. If multiagent systems are working for weeks, how does the human meaningfully approve what their agent can do? Agentic platform features like auto approval for tool calls helps, but there's deeper permissioning needed to get this right at high scale. And then we have agents working with their own identity, which is even less solved, and how we apply that to agents working together - how does an agent trust another agent that asked it for something?
身份与访问控制
智能体认证(authentication)本身就有不少要解决的问题——识别出到底是哪个智能体做了某个动作。这个领域可以直接受益于微服务世界里已经打磨得很硬的那些构件,比如用短期证书而不是静态密钥的工作负载身份;大家正开始把这些概念套用到智能体上。
而智能体授权(authorization)作为一个问题空间就要开放得多——你愿意让一个智能体做哪些事?许多智能体是代表某个人在工作,用的是那个人的身份和凭证去访问各种系统、执行各种动作。这里本身就有一个问题:如何识别一次动作到底是智能体做的,还是本人做的。但这种"代为处理"(on-behalf-of,OBO)设置里还有更深的问题,比如同意(consent)。如果一个多智能体系统持续工作好几周,一个人要怎样"有意义地"去批准他的智能体可以做什么?"工具调用自动批准"这样的 agentic 平台能力有帮助,但要在高规模下把这件事做对,还需要更深的权限体系。再往下,还有以自身身份工作的智能体——这块更是远远没有解决,而且当多个智能体协作时还要延伸出一个问题:当另一个智能体来请求做某件事时,这个智能体凭什么去信任对方?
Security and isolation
Over the past few months and especially in the past couple weeks, we've seen vulnerabilities surfacing faster than the industry can patch them, much of it in critical core software that many systems rely on. Now think about thousands of persistent agents running on top of that core software, writing untrusted code and reading untrusted content. Those unpatched vulns get a lot more dangerous in that world, and one of the most important things for software engineers to do is get creative about how to harden that software - kernels, browsers, runtimes, and common libraries.
Multiagent systems add another twist. A compromised agent could pass injected instructions to other agents it's working with, and those instructions could end up in memory for future agents to accidentally trust. Blast radius gets real - how much of the multiagent system is at risk when that first agent got compromised? Isolation becomes important in that world - agents should use strongly isolated sandboxed execution environments to take actions. I wrote a bit about that here.
安全与隔离
过去几个月,尤其是最近几周,我们已经看到漏洞冒出来的速度比行业打补丁的速度还快,而且很多都是在许多系统赖以运行的核心软件里。现在再去想象:成千上万个长期运行的智能体跑在这些核心软件之上,一边写着不可信的代码,一边读着不可信的内容。那些没打补丁的漏洞在这种世界里会危险得多,而软件工程师要做的最重要的一件事,就是在"如何让这些软件更坚固"上发挥创造力——内核、浏览器、运行时和常用库都在其中。
多智能体系统还添加了一层新花样。一个被攻陷的智能体可以把注入的指令传给与它协作的其他智能体,而这些指令可能进入"记忆"里,被后续的智能体不加防备地信任。爆炸半径(blast radius)于是变得非常真实——当最初那个智能体被攻陷时,整个多智能体系统有多少部分会随之暴露在风险之中?在这种世界里,隔离就变得很重要——智能体执行动作时,应当使用强隔离的沙箱执行环境。关于这一点我写过一点内容,在这里。
Monitoring and halting
Lastly, we need the tools to make sure these multiagent systems are working in a positive direction, and that we catch them and cut them off if they're not. We need OTel tracing across agents, and to ensure that an agent can never modify its own trace or another agent's trace. We need kill switches that monitoring can trigger, that truly stop the whole system at once and cut off its creds. Really the hard part is how to know when the system is going in the wrong direction. Well-intentioned persistence to solve a problem can look a lot like maliciousness - repeated attempts to reach a domain that isn't on its allowlist, for example.
监控与停机
最后,我们需要工具去确保这些多智能体系统朝着正向的方向运转,并在它们偏离时抓到它们、把它们切掉。我们需要跨智能体的 OTel 链路追踪(tracing),并保证任何一个智能体都不能修改自己的 trace 或别人的 trace。我们需要监控能够触发的 kill switch,能真正"一下子"停住整个系统并切断它的凭证。真正难的部分是如何判断系统正在朝错误的方向走——"出于善意的持续尝试去解决一个问题"看起来可以非常像恶意行为,比如反复尝试访问一个不在白名单上的域名。
This is for the engineers feeling an identity crisis because they think their knowledge and skills won't be important - you and your skills can be the difference in whether or not these problems get tackled and multiagent systems can solve the world's hardest problems. You can and should go deep in these domains, learn new things, and form opinions. So much of this problem space is wide open for new solutions. It's an incredibly exciting time to put your skills to work.
这段话写给那些正陷在身份危机里、觉得自己手上的知识和技能不再重要的工程师——你和你的技能,恰恰可能决定这些问题到底会不会被解决、多智能体系统到底能不能去解决世界上最难的问题。你完全可以,也应该,在这些领域里深入下去、学新的东西、形成自己的观点。这个问题空间里有太多地方对新方案是完全开放的。此刻,正是把你的技能用起来的、极其激动人心的时刻。
Katelyn Lesse — head of platform @anthropicai. previously @stripe @betterment. lfgm.
Katelyn Lesse——Anthropic 平台负责人。此前就职于 Stripe、Betterment。lfgm.

评论
评论由 GitHub Discussions 提供。