Share

The internet was built on a principle so elegant it reads like philosophy: the network should not understand what it carries. In their 1984 paper, Saltzer, Reed, and Clark argued that functions like reliability, security, and correctness "can completely and correctly be implemented only with the knowledge and help of the application standing at the endpoints."¹ The network moves packets. Everything else belongs to the edges.

For the engineers who operated ISP infrastructure in the 1990s, this wasn't an abstraction. It was a daily operational posture. They routed traffic based on source IPs, destination IPs, and port numbers. They could see where packets were going. They could not see, and were architecturally discouraged from caring about, what those packets were doing when they got there.

Now autonomous AI agents send HTTP requests through the same architecture, executing multi-step commercial transactions across infrastructure that still treats each packet in isolation. In July 2026, Cloudflare announced controls to categorize and block automated AI traffic by inferred purpose, an infrastructure layer attempting to answer the very question the end-to-end principle said the network couldn't answer.

We spoke with Gil Hopsworth, who ran network operations at a regional ISP in North Carolina's Research Triangle from 1994 to 2002 and now teaches networking at a community college nearby. He occupies that particular liminal space between people who remember what the early commercial internet smelled like (the ozone from overheating modems) and students who think "BGP" is a K-pop group. To the extent that Gil is a composite of several engineers we've spoken with, and to the extent that his surname contains a networking pun he insists is coincidental, we present this conversation as reconstructed.

You operated network infrastructure during the period when the internet was commercializing. What did the end-to-end principle feel like from inside a NOC?

Gil: It felt like being a postal worker who's been told, very firmly, that opening the mail is a federal crime. Except in our case it wasn't a crime. It was a philosophy. We had routers, we had BGP tables, we had port numbers. Port 80 meant HTTP, port 25 meant email. That's what we knew. That's what we were supposed to know.

I could tell you that a packet was going from a Raleigh subnet to a server at MIT. I could not tell you whether that packet was a grad student submitting a thesis or someone buying a first edition on some early web storefront. And the thing is, this was considered correct. We weren't failing to observe something. We were succeeding at not observing it.

David Clark, one of the end-to-end principle's architects, wrote in 1988 that the internet's design goals were explicitly prioritized: survivability first, accountability last.² Did that priority ordering feel deliberate to you at the time?

Gil: Oh, we didn't sit around reading Clark papers. We were racking Cisco 2500s and arguing about whether our T-3 could handle the load from AOL users discovering the web. But the consequences of that ordering were everywhere. We had no mechanism to account for what traffic was doing. BGP told us reachability, how to get packets from here to there. It had, as someone put it, "limited expressive power."³ It described where, never why.

And honestly? That felt fine. Because at the other end of every packet was a person. A human being sitting at a keyboard, or at worst a human being's server running a script they'd written. If something went wrong, if a transaction failed, if an email bounced, if someone accessed something they shouldn't have, there was a human in the loop who could pick up a phone.

“

The accountability wasn't in the network. It was in the chair.

That's a striking image.

Gil: And nobody questioned it because it worked. The 1984 paper actually uses the example of airline reservations. It says "an agent can be relied upon to keep trying, through system crashes and delays, until a reservation is either confirmed or refused."¹ They meant a human travel agent. A person with a phone and a Rolodex and professional liability. The network didn't need to guarantee the reservation went through because Sandra at the travel desk would keep trying until it did.

That word "agent" in the original paper is doing a lot of work these days.

When did you first notice the network might need to understand more about what it carried?

Gil: Ninety-seven, ninety-eight. Spam was the first crack. When you're routing ten million emails a day and a growing percentage of them are fraudulent, the purity of "we just move packets" starts feeling less like a principle and more like an excuse. Then Napster hit and suddenly our upstream providers cared very much about what was inside those packets on port 80.

Here's the thing people forget. Cisco had deep packet inspection capability from '91.⁴ The technology to look inside existed for years before the commercial pressure to use it became overwhelming. We chose not to look. That was a choice, grounded in an architectural philosophy. Not ignorance. Restraint.

So the deliberate ignorance was a feature, and then it became a liability.

Gil: That's exactly backward, actually. It was always both. The RFC from '96, the IAB one that codified all this, says explicitly: "the intelligence is end to end rather than hidden in the network" and "endpoints should not depend on the confidentiality or integrity of the carriers."⁵ The network was supposed to be dumb. The endpoints were supposed to be smart. And for a while, "smart" meant "operated by a human who can be held responsible."

The liability was always there. We just had humans papering over it.

And now the endpoint might be an AI agent executing a purchase.

Gil: Yeah. And this is where I get... look, I teach eighteen-year-olds about TCP/IP. They're going to build systems where agents negotiate contracts, book travel, move money. And the infrastructure those agents operate on was designed, explicitly designed, around the assumption that correctness and accountability live at the endpoints.

Clark wrote that "the gateway cannot directly see the existence of this sequence, because it is forced to deal with each packet in isolation."² Think about what that means for a multi-step purchase. Search, select, cart, payment, confirm. That's a sequence. The network sees five separate packets. It has no concept that they constitute a commercial commitment.

Does that worry you?

Gil: What worries me is that people think the solution is making the network smarter. Adding more inspection, more classification, more inference about what traffic is for. Categorize the traffic, block what you don't like. But the 1984 paper was right about something fundamental: the network cannot verify end-to-end correctness because it doesn't have enough context.¹ An intermediate node can see that an agent sent an HTTP POST. It cannot see whether that POST was authorized by a human, within a spending limit, consistent with a mandate. That information only exists at the endpoints.

The problem is that the endpoints have changed. The endpoint used to be Sandra at the travel desk. Now it might be an agent that was told "book something reasonable" and interpreted that however its model interpreted it.

The principle still holds. Correctness lives at the edges. But the edges need to be a lot more articulate about what they're doing and on whose authority.

You've been thinking about this for a while.

Gil: I've been routing packets for thirty years. For twenty-five of those years, I didn't need to think about it. The humans handled accountability and the network handled delivery and everyone was fine.

Now the humans are stepping back from the keyboard and the network is still doing exactly what it was designed to do: moving packets without understanding them.

“

The architecture is working perfectly. That might be the problem.

J.H. Saltzer, D.P. Reed, D.D. Clark, "End-to-End Arguments in System Design," ACM Transactions on Computer Systems, Vol. 2, No. 4, November 1984. https://web.mit.edu/saltzer/www/publications/endtoend/endtoend.pdf ↩ ↩² ↩³
David D. Clark, "The Design Philosophy of the DARPA Internet Protocols," ACM SIGCOMM Computer Communication Review, Vol. 18, No. 4, August 1988. https://www.cs.princeton.edu/courses/archive/fall06/cos561/papers/clark88.pdf ↩ ↩²
Referenced in "Making the Unaccountable Internet: The Changing Meaning of Accounting in the Early ARPANET." https://arxiv.org/pdf/2201.11884 ↩
Centre for Internet and Society, "Deep Packet Inspection: How it Works and its Impact on Privacy." https://cis-india.org/internet-governance/blog/deep-packet-inspection-how-it-works-and-its-impact-on-privacy ↩
Brian Carpenter (ed.), "Architectural Principles of the Internet," RFC 1958, Internet Architecture Board, June 1996. https://datatracker.ietf.org/doc/html/rfc1958 ↩

You operated network infrastructure during the period when the internet was commercializing. What did the end-to-end principle feel like from inside a NOC?

David Clark, one of the end-to-end principle's architects, wrote in 1988 that the internet's design goals were explicitly prioritized: survivability first, accountability last.² Did that priority ordering feel deliberate to you at the time?

“

The accountability wasn't in the network. It was in the chair.

That's a striking image.

That word "agent" in the original paper is doing a lot of work these days.

When did you first notice the network might need to understand more about what it carried?

So the deliberate ignorance was a feature, and then it became a liability.

The liability was always there. We just had humans papering over it.

And now the endpoint might be an AI agent executing a purchase.

Does that worry you?

The principle still holds. Correctness lives at the edges. But the edges need to be a lot more articulate about what they're doing and on whose authority.

You've been thinking about this for a while.

Now the humans are stepping back from the keyboard and the network is still doing exactly what it was designed to do: moving packets without understanding them.

“

The architecture is working perfectly. That might be the problem.

J.H. Saltzer, D.P. Reed, D.D. Clark, "End-to-End Arguments in System Design," ACM Transactions on Computer Systems, Vol. 2, No. 4, November 1984. https://web.mit.edu/saltzer/www/publications/endtoend/endtoend.pdf ↩ ↩² ↩³
David D. Clark, "The Design Philosophy of the DARPA Internet Protocols," ACM SIGCOMM Computer Communication Review, Vol. 18, No. 4, August 1988. https://www.cs.princeton.edu/courses/archive/fall06/cos561/papers/clark88.pdf ↩ ↩²
Referenced in "Making the Unaccountable Internet: The Changing Meaning of Accounting in the Early ARPANET." https://arxiv.org/pdf/2201.11884 ↩
Centre for Internet and Society, "Deep Packet Inspection: How it Works and its Impact on Privacy." https://cis-india.org/internet-governance/blog/deep-packet-inspection-how-it-works-and-its-impact-on-privacy ↩
Brian Carpenter (ed.), "Architectural Principles of the Internet," RFC 1958, Internet Architecture Board, June 1996. https://datatracker.ietf.org/doc/html/rfc1958 ↩

Gil Hopsworth on Routing Packets for Thirty Years Without Asking Why

You operated network infrastructure during the period when the internet was commercializing. What did the end-to-end principle feel like from inside a NOC?

David Clark, one of the end-to-end principle's architects, wrote in 1988 that the internet's design goals were explicitly prioritized: survivability first, accountability last.² Did that priority ordering feel deliberate to you at the time?

That's a striking image.

When did you first notice the network might need to understand more about what it carried?

So the deliberate ignorance was a feature, and then it became a liability.

And now the endpoint might be an AI agent executing a purchase.

Does that worry you?

You've been thinking about this for a while.

Gil Hopsworth on Routing Packets for Thirty Years Without Asking Why

You operated network infrastructure during the period when the internet was commercializing. What did the end-to-end principle feel like from inside a NOC?

David Clark, one of the end-to-end principle's architects, wrote in 1988 that the internet's design goals were explicitly prioritized: survivability first, accountability last.² Did that priority ordering feel deliberate to you at the time?

That's a striking image.

When did you first notice the network might need to understand more about what it carried?

So the deliberate ignorance was a feature, and then it became a liability.

And now the endpoint might be an AI agent executing a purchase.

Does that worry you?

You've been thinking about this for a while.

Gil Hopsworth on Routing Packets for Thirty Years Without Asking Why

You operated network infrastructure during the period when the internet was commercializing. What did the end-to-end principle feel like from inside a NOC?

David Clark, one of the end-to-end principle's architects, wrote in 1988 that the internet's design goals were explicitly prioritized: survivability first, accountability last.2 Did that priority ordering feel deliberate to you at the time?

That's a striking image.

When did you first notice the network might need to understand more about what it carried?

So the deliberate ignorance was a feature, and then it became a liability.

And now the endpoint might be an AI agent executing a purchase.

Does that worry you?

You've been thinking about this for a while.

Footnotes

Gil Hopsworth on Routing Packets for Thirty Years Without Asking Why

You operated network infrastructure during the period when the internet was commercializing. What did the end-to-end principle feel like from inside a NOC?

David Clark, one of the end-to-end principle's architects, wrote in 1988 that the internet's design goals were explicitly prioritized: survivability first, accountability last.2 Did that priority ordering feel deliberate to you at the time?

That's a striking image.

When did you first notice the network might need to understand more about what it carried?

So the deliberate ignorance was a feature, and then it became a liability.

And now the endpoint might be an AI agent executing a purchase.

Does that worry you?

You've been thinking about this for a while.

Footnotes

David Clark, one of the end-to-end principle's architects, wrote in 1988 that the internet's design goals were explicitly prioritized: survivability first, accountability last.² Did that priority ordering feel deliberate to you at the time?

David Clark, one of the end-to-end principle's architects, wrote in 1988 that the internet's design goals were explicitly prioritized: survivability first, accountability last.² Did that priority ordering feel deliberate to you at the time?