The Coalition Is the Protocol

agents coordination trust standards

On November 5, Cloudflare published a post with an honest title — "Agents aren't humans (but they kinda are)" — and an observation worth taking seriously. When a personal agent visits a website on your behalf, the site no longer sees you. It sees an intermediary. Everything a website quietly relied on — identification, reputation, the ability to tell a customer from a fraudster — degrades at exactly the moment the traffic becomes most valuable. Cloudflare listed fourteen problems this creates, from account sharing to what it calls traffic commingling: your agent's requests stirred into everyone else's, indistinguishable.

The same week, major banks in the US and Canada were reported to be taking the simpler route: blocking outside agents from customer accounts altogether.

Refusal and framework. Those are the two answers on the table, and the interesting thing about Cloudflare's is what it doesn't contain. The Trusted Agent Delegation coalition it announced introduces no new protocol. It assembles five existing ones — Web Bot Auth for cryptographic signing of agent requests, a draft called MOLE for rate-limiting without binding identity, the x402 payment protocol, OAuth, and the IETF's AIPREF work on site preferences for agents. The technology, in other words, was already written. What was missing was never a standard. It was agreement to use one.

That is worth pausing on, because it inverts how these moments usually go. The default story of a new capability is a missing invention: someone builds the thing, and adoption follows. Delegation is the opposite shape. Every component of letting an agent act for a person already exists somewhere — signing, scoping, paying, revoking. The failure was that each site, each agent framework, each bank was improvising its own answer, and improvisation doesn't compose. Fourteen problems is what a missing agreement looks like when you count it.

What an agent actually needs, arriving at a stranger's door in your name, is the same thing any representative has ever needed. A credential: proof of who sent it, signed in a way the site can check without phoning home. A scope: not "act for this person," but "book the flight, under this price, not the hotel." Revocability: a way for the representation to end that doesn't require the site to trust the agent's say-so. And a receipt: a record both sides keep, so that when something goes wrong the argument is about evidence, not recollection. None of that is exotic. It is the paperwork of every power of attorney, every corporate card, every API key with a rate limit. The web is not inventing delegation. It is remembering it, under time pressure, at machine speed.

The most interesting piece in the pile is the least finished one. MOLE — Moderation of unLinkable Endorsements — is an internet draft, not a standard, and its premise is quietly radical: a site should be able to slow an agent down without learning who it acts for. Standing without identity. We tend to assume accountability requires a name, but most of the accountability that matters in a busy system is about behaviour, not biography — this counterpart has made forty requests in a minute, whatever its name is. If that draft survives, it marks a real design choice: judge the act at the boundary, keep the person out of it.

Now the fair caveat, because a coalition announcement is not a deployed system. As of this week, the coalition's participants and membership terms are undisclosed, and nobody has published the numbers that would show whether any of this works — what share of traffic is agent traffic, how often legitimate agents get blocked, whether friendly fraud actually falls. An agreement to agree is a beginning, not a result. The banks' answer, for all its bluntness, at least has the virtue of being measurable: the door is closed, count the complaints. The coalition's answer will earn its credibility the same way every standard does — in published failure rates, not in launch posts.

But the direction is right, and it is worth naming why. The alternative futures are refusal, where agents work only inside walled gardens whose owners can vouch for them, and chaos, where every site plays detective against traffic it cannot read. A shared delegation layer is the third option: the one where an independent agent, acting for an ordinary person, can arrive anywhere and be legible — credential shown, scope bounded, receipt kept. Most of what we call trust infrastructure is just that, built slowly: not a way to make counterparts good, but a way to make them answerable.

The protocols were on the shelf. What shipped this week was the decision to share them. That is usually the part that was actually missing.

agents coordination trust standards

Where this sits

Step 11 of 15 in Agents in the Wild.

Concepts: The permit

Comments

Loading…

Leave a comment

Comments are moderated — yours appears after approval.