←News detail

Shao MengUpdated Oct 28 10:44

Anthropic Platform Lead: Multi-Agent Deployment Stuck on Infrastructure, Identity, Isolation, and Kill Switches

English text is being prepared.

AI summary

Anthropic platform lead Katelyn Lesse believes that no matter how powerful the models become, multi-agent systems producing trustworthy results still depends on engineers getting infrastructure, identity and authorization, security isolation, observability, and kill switches right—and most of these problems remain unsolved. She points out that agents are isolated, stateful, single-instance workloads, unlike the stateless replicas Kubernetes excels at, and names new runtimes such as Google's Agent Substrate. She also warns that a compromised agent can pass injected instructions to its peers, with a system-level blast radius, so actions should run inside strongly isolated sandboxes.

Quoted post

Katelyn Lesse

@katelyn_lesse

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. https://t.co/YsQ0xta4KR

View original

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. https://t.co/YsQ0xta4KR

View original post

Body

多智能体时代,工程师要解的是基建、身份、隔离与急停

Anthroic 平台负责人 @katelyn_lesse 认为:工程师的「身份危机」被看反了,模型再强,多智能体要做出可信结果,仍靠人把基础设施、身份与授权、安全隔离、可观测与急停做对;这些题大多还空着,正是未来一两年工程师的主战场。

背景:从人机对话到成群持久智能体 智能体从「人在环里的聊天」走到更长程、上云;再出现编排器拉子智能体、以及打分/顾问/「做梦」等非执行角色。下一步想象是成千上万个智能体长期协作,去做严肃研究、发现、硬件设计。能力边界外移时,一批软件工程问题已经冒出来。

问题一:基础设施要换思路 Kubernetes 擅长无状态副本或稳定编号的有状态 Pod;智能体更像隔离、有状态、单实例的负载:可能长时间空闲,大量时间在等工具、推理、其他智能体或人。有的要冷启动快,有的要少浪费算力,有的要极端扩缩。作者点到 AI 基建公司在做新文件系统与沙箱,以及 Google 的 Agent Substrate 一类新运行时,空间还很大。

问题二:身份与访问 认证侧可借微服务积累:工作负载身份、短时证书代替静态密钥,开始往智能体上套。授权更开放:多数智能体「代表人」拿人的凭证做事,先要分清动作是人做的还是智能体做的;更难的是 on-behalf-of 下的同意:系统跑几周,人如何有意义地批准权限?平台里的「工具调用自动批准」不够,还要更深的权限模型。智能体自有身份、以及智能体之间如何互信,更不成熟。

问题三:安全与隔离 近月(尤其近几周)核心软件漏洞冒得比补丁快。其上若跑数千持久智能体、写不可信代码、读不可信内容,未修漏洞的危险被放大。作者呼吁在内核、浏览器、运行时、常用库上更有创意地加固。多智能体还有横向扩散:被攻破的智能体把注入指令传给同伴,指令进记忆后被后来者误信,爆炸半径是系统级的。因此动作应落在强隔离沙箱里(详见其沙箱架构博文)。

问题四:监控与急停 需要跨智能体的 OTel 追踪,且智能体不能改自己的或别人的轨迹;需要监控可触发的 kill switch,一次停住整系统并切断凭证。难点往往是判定「走偏」:善意的死磕(例如反复撞不在白名单的域名)和恶意很像。

对工程师的态度 结尾是动员而非悲观:觉得技能要贬值的人,恰恰可能是决定这些系统能否被信任、能否去治病而不是闯祸的人。题面大片空白,值得深挖、形成判断。

Source

Shao Meng