Research August 2026 9 min GDA Research

01 The output of a protocol business is adoption. Software is the input, and the two are routinely confused by the people funding it.

02 The diagnostic is one question: can an organisation with no relationship to the author implement this without speaking to them, and be understood by a second organisation that also has none?

03 Until that has happened once, an estate implementing its own specification is a demonstration, however many properties it spans.

04 Neutral governance raises the value of a standard rather than diluting it, which is why the recent agent protocols were donated to foundations by the companies that wrote them.

05 The commercial position is the best implementation and the network around the standard, not ownership of the standard.

06 The worst-timed exit in this category is the one taken just before the count begins to compound.

An infrastructure business that publishes a specification has two visible outputs, and they are easy to mistake for one another. There is the software — repositories, services, documentation, a reference implementation — and there is adoption, meaning the count of organisations that have implemented the specification and now depend on it.

Only the second is the asset. The software is the cost of producing it. This is not a subtle distinction, but it is obscured in practice, because software is visible, measurable and gratifying to produce while adoption is slow, external and largely outside the author's control.

The consequence for anyone assessing such a business is that engineering output is close to uninformative. A specification implemented by nobody is a document, and a comprehensive one is a longer document. The question is never how much has been built. It is who else has built against it.

There is a specific and common failure mode: an author writes a specification, implements it faithfully across every property they control, and produces something that works, is internally consistent, and has never once been tested by a party who could have said no.

This is not wasted work. Implementing your own specification several times is how it stops being theoretical, and an author who has not done it is publishing an untested idea. But it must not be mistaken for adoption, and it frequently is — particularly where the author controls a family of businesses, because the resulting count looks like a network.

The tell is whether any implementer could have refused. Inside one ownership structure nobody can, so the fact that they did not carries no information. A specification adopted across nine properties by one principal has been adopted once.

The question that resolves this is narrow and answerable. Can an organisation with no relationship to the author discover the specification, implement it against their own systems without speaking to the author, and be understood by a second organisation that also has no relationship to the author?

Each clause is doing work. "No relationship" removes the ownership effect. "Without speaking to the author" tests whether the documentation is genuinely sufficient, which almost no specification's authors can judge from the inside. "Understood by a second organisation" is what separates a standard from a well-documented product, because a format only one party reads is an integration.

The first time all three clauses hold at once, something has changed that no additional engineering could have produced. Before that moment the business has a specification. After it, the business has evidence.

The count of independent implementers is not a smooth measure of progress. It changes meaning at roughly each order of magnitude, and the transitions matter more than the increments.

At one, the specification is sufficient to be implemented by a stranger. That is a fact about the documentation as much as the design, and it is the hardest single step.

At ten, the first was not an anomaly. The pattern of who adopts and why becomes readable, and this is usually where the specification needs its first backward-compatible correction.

At one hundred, third parties begin building tools for the specification rather than for the author's implementation of it. The author stops being the only source of gravity.

At one thousand, the format is a fact of the market. Vendors implement it because their customers assume it, and the author's commercial position is set by the quality of what they built around it rather than by authorship.

Above that, the question facing a large platform is no longer whether the layer matters, but whether to partner, interoperate, compete or acquire — and that question gets asked without the author in the room.

The counterintuitive finding, and the one most often resisted by owners, is that a standard becomes more valuable to its author when the author visibly cannot control it.

The mechanism is straightforward. An organisation deciding whether to build against a format is underwriting a dependency, and the risk being priced is that the owner changes it, prices it, or withdraws it once the implementer is committed. Neutral governance removes that risk, and removing it is worth more in adoption than control is worth in optionality.

This is now an observed pattern rather than a theory. The major open agent protocols were donated by the companies that authored them into vendor-neutral foundations, and the largest platform companies have signed on to governance they do not control. They did not do this out of generosity. They did it because a protocol one vendor owns is a product, and they were not trying to sell a product.

The commercial consequence for an author is that the money is not in the standard. It is in being the best implementation of it, holding the operating network built around it, and being the party with the longest experience of running it in production.

This paper does not price outcomes and declines to publish a range for any of them, because the honest input — the independent implementer count — is usually near zero at the point the question is first asked, and a valuation built on a projected adoption curve is a projection wearing a valuation's clothes.

What can be stated is the shape of the error. The distribution of outcomes for a protocol business is wide and its mass sits late, because value accrues after adoption compounds rather than before. An acquirer approaching such a business early is buying the option on that compounding at a price set before it is visible.

That makes the worst-timed exit in this category a specific and identifiable one: the sale taken after the specification is complete and before the escape test has been passed repeatedly. At that point the seller has absorbed the full cost of building the thing and transferred the entire value of its being adopted.

The corresponding discipline is easy to state and hard to hold. Manage toward the count, not toward the conversation. An owner who can show a rising number of independent implementers is negotiating from evidence; one who can show only a larger codebase is negotiating from a story.

Three questions separate a protocol business from a well-documented product, and none of them is about the technology.

First: is the count published? An author who reports independent implementers, including when the number is zero, is measuring the thing that matters. An author who reports repositories, features or properties is measuring the input and hoping it reads as the output.

Second: could an implementer have refused? Where every adopter shares the author's ownership, the count is one regardless of what it says.

Third: what happens if the author stops? A specification that continues to function, be implemented and be governed without its originator has become infrastructure. One that does not was always a product with unusually good documentation — which can be an excellent business, and should be valued as one.

0 A specification. However complete, it has not been tested by anyone who could refuse.

1 The documentation is sufficient for a stranger. The hardest step, and rarely the one celebrated.

10 Not an anomaly. The adoption pattern becomes readable and the first correction is usually needed.

100 Third parties build tools for the format rather than for the author's implementation of it.

1,000 A fact of the market. Commercial position is decided by what was built around it, not by authorship.

10,000+ Large platforms decide whether to partner, interoperate, compete or acquire — without the author present.