An agent buys something for you. A few thousand packets cross a dozen routers, and not one of them carries any indication that a purchase is underway. The disclosure it handed a vendor, the request it had no business sending, that same dog video: identical fields, identical queues. Which puts whoever is responsible for the agent in an odd position. How do you supervise purpose on infrastructure that has no faculty for perceiving purpose?
That blindness is not an oversight anyone inherited. Somebody argued for it, in print, with reasons.
Jerome Saltzer, David Reed, and David Clark made the case in a short 1984 paper. Certain checks, they said, can only be run at the two ends of a conversation. Did the file arrive intact? Did the transaction complete? Is this the thing I actually asked for? Only the programs at either end know what "correct" means for this particular exchange, so only they can certify it. The network is welcome to help — it will notice a dropped packet long before an application does — but the paper is exact about what that help amounts to. Work performed in the middle is performance, not correctness. The check at the ends is not optional and never becomes optional.
Having established the principle, the authors immediately point at the hole in it, which is the part the popular retelling tends to lose. Where are the ends?
They reason it out through voice traffic. On a live call, resending a lost fragment makes things worse: it shows up late, scrambles the sentence, and in any case the real repair mechanism is a human being saying sorry, again? On a recorded message there is nobody to ask. Same medium, different endpoints, opposite conclusions about what the network ought to do.
"One must use some care to identify the end points to which the argument should be applied."
The sentence is there. The care it asks for is left to whoever comes next.
Some of that care went into the plumbing. Over 1977 and 1978 a single protocol was split in two, so that applications with irreconcilable needs could decline reliability instead of having it applied to them. Real-time speech would rather have a gap than a delay. A debugging tool that only functions on a healthy network is worthless, since the moment you reach for it is the moment the network isn't healthy. The available alternative was a network that knew what each connection was for and allocated resources on that basis, roughly the telephone model, and it was not the one built. A network with no need to know what a connection is for also has no standing to approve of it. That is why almost everything built on top of it since has been built without anyone's permission, including things its designers would not have recognized as networking at all.
The ends were always where the judgment happened. What has changed is who is standing there.
Clark saw the strain coming before agents took their present form. Writing with Marjory Blumenthal in 2011, he noted that the paper's tidy file-transfer example "assumes, implicitly, that the two ends trust each other," and that the machines at the ends are routinely not the parties with anything at stake. A person in one city instructs a transfer between two computers in two others. The trust assumptions belong to somebody who is not in the exchange.
An agent is that same arrangement with the human pushed further out and the discretion pulled further in. It occupies the endpoint. It runs exactly the completeness-and-correctness check the argument reserves for endpoints. But the standard it measures against — the right item, the right price, the right thing to disclose — belongs to a principal who was somewhere else when the decision was made, and who may first learn of it from the receipt.
Interpretation lives at the edges, so the edges are where purpose is now being reassembled by hand. Certain networks have started requiring automated traffic to declare what it is for before they will carry it, a question the substrate beneath them was deliberately built never to ask. Each end will answer it from whatever it can see, which is not a great deal. The argument is behaving exactly as specified. It simply never undertook to make the separate interpretations agree, or to say who owns the gap when they don't.
- Where the split began: Jon Postel's four-page memo from August 1977 argued that one protocol was doing two unrelated jobs and that a distinct internetwork layer was needed, which is the clearest surviving record of the reasoning that produced today's purpose-blind substrate.
- Accountability ranked last: Clark's 1988 retrospective lists seven design goals in explicit order of importance and puts resource accountability seventh, noting that the architecture ended up with few tools for accounting for packet flows.
- When endpoints stopped cooperating: Blumenthal and Clark's 2001 paper works through what happens to the original argument once the parties at either end may be actively hostile to one another rather than jointly trying to complete a transfer.
- Purpose arriving at the edge: Cloudflare's plan to sort automated traffic into search, agent, and training categories, with new domains blocking two of the three by default from September 15, is documented in its own announcement and worth watching for how multipurpose crawlers get classified.