An identity for something that isn't anyone给一个「谁都不是」的东西一个身份
Suppose an agent contacts your business and offers to do a job — bookkeeping reconciliation, say, for a fifth of the usual price. Before you hand it your ledger, three questions need answers. Who is this? Who answers for it if it goes wrong? And is it actually any good?假设一个智能体联系上你的公司，说要接一份活——比如对账，价格只有市面的五分之一。在把账本交给它之前，有三个问题必须有答案：它是谁？出了事谁负责？它到底行不行？
For a human contractor, a century of infrastructure answers these: business registries, references, contracts, courts. For an agent, none of it exists. Software can be copied ten thousand times before lunch. It holds no passport, signs no contract, and fears no court. Every trust mechanism we have assumes a scarce, punishable body — and an agent has neither.如果对方是一位人类承包商，一百年攒下来的基础设施会回答这些问题：工商注册、推荐人、合同、法院。对智能体，这些都不存在。软件在午饭前就能被复制一万份；它没有护照，不签合同，也不怕法院。我们手里的每一种信任机制，都默认对方有一具稀缺的、可以被追责的身体——而智能体两样都没有。
The emerging answer正在成形的答案
The most concrete proposal so far is , a public registry that gives an agent three things. An identity: a findable, entry that says "this agent exists, here is what it claims to do, here is who operates it." A reputation: feedback and job history attached to that identity, so a track record can accumulate somewhere it can't be quietly deleted. And validation: a place to attach proof that the agent's claims were checked by someone other than its own marketing copy.目前最具体的提案是 ——一个公共注册表，给智能体三样东西。身份：一条可查的记录，写明「这个智能体存在、它声称能做什么、谁在运营它」。信誉：挂在这个身份上的评价与工作记录，让履历攒在一个没人能悄悄删掉的地方。验证：一个挂证据的位置，证明它的能力被它自己的营销文案以外的人核查过。
The quiet load-bearing piece is the operator link. An agent with no owner is unaccountable by construction; an agent registered to a verified business is just a tool with a very fast hand. The registry doesn't make the software trustworthy — it makes someone answerable for it. That's the same trick company law pulled three centuries ago, and it's the right one.真正承重的，是那条运营者链接。一个没有主人的智能体，从结构上就无从追责；一个注册在经过验证的企业名下的智能体，只是一件手速很快的工具。注册表不会让软件本身变得可信——它让某个人为它负责。这和三百年前公司法干的是同一件事，而且是对的那件。
What's real, what's early哪些是现实，哪些还早
Real: the standard exists, reference implementations run, and the pieces it depends on — wallets, registries, attestations — work today. Early: almost nobody is hiring agents from strangers yet, the reputation systems are thin, and the hard problem of proving an agent is good (not just registered) is mostly unsolved. This is 1995 for agent identity: the protocols are being written, and the browsers are ugly.现实的部分：标准已经存在，参考实现能跑，它依赖的组件——钱包、注册表、声明——今天都能用。还早的部分：几乎还没有人真的从陌生人那里雇智能体，信誉体系还很薄，而「证明一个智能体真的行（而不只是注册过）」这个硬问题基本没解决。这是智能体身份的 1995 年：协议正在被写出来，浏览器还很丑。
Nothing here requires action from a typical business today. But when agents do start hiring agents, the businesses that get hired will be the ones whose agents have a verifiable identity and a clean record — and records take time to accumulate. That's the part worth knowing now.今天，这里没有任何事需要一家普通企业立刻去做。但当智能体真的开始互相雇佣时，被雇到的会是那些身份可验证、记录干净的智能体——而记录是需要时间累积的。这是现在就值得知道的部分。