When Philipp Hölke wrote that in agent economics distribution happens in the routing table, not the browser, he was describing a market structure that, by his own admission, does not exist yet. This is an audit of how much of it exists in July 2026. The short answer: discovery is real, selection is real but crude, and the routing table itself, the neutral layer that compares vendors on contract, price, and track record per invocation, is still missing. Which means the most valuable position in the agent economy is currently unoccupied.
Start with what "routing" would have to mean, because the word hides four separate jobs. First, discovery: the agent has to know your product exists. Second, evaluation: it has to decide whether your product can do the job to specification. Third, transaction: it has to invoke you, and increasingly, pay you. Fourth, feedback: the outcome has to update what the agent chooses next time. SaaS distribution bundled all four into one mechanism, a human noticing your brand, and called it marketing. The agent economy is unbundling them into infrastructure, one layer at a time and out of order.
Discovery is the furthest along. The MCP Registry launched in September 2025 as an open catalog, an app store for MCP servers, and it is deliberately metadata-first: it points at the code and feeds downstream marketplaces, including private enterprise sub-registries that mirror the public one. Around it has grown a fractured bazaar, at least five distinct channel types by one survey's count: open protocol registries, developer catalogs, consumer assistant stores, suite-embedded directories, and the cloud procurement marketplaces, where AWS, Google, and Microsoft now list agents and MCP servers next to virtual machines. Fractured, but functional. If you ship a capability today, an agent can find it. This is the yellow pages phase: comprehensive, unranked, and quality-blind.
Evaluation is where it gets interesting, because this is the layer that actually moves usage, and almost nobody building products understands how it works yet. Getting connected is not getting chosen. A server can sit in an agent's tool list for months and never be invoked, the agent-economy equivalent of shelf space with no sales. When an agent decides which of its available tools to call, the mechanics are unglamorous: it reads your tool descriptions and schemas, and it matches them, semantically, against the user's intent.
Which produces a strange new discipline, a kind of SEO where the searcher has read everything and clicks nothing out of habit. The empirical findings from people who watch these selection traces are wonderfully concrete. Tools whose descriptions lead with verbs matching the user's phrasing get picked. Tools that declare their output shape, returns a JSON object with these fields, get picked over tools that promise "the result," because the agent can plan what it will do with the response before spending the call. Tools whose schemas require something the agent cannot supply mid-flight, an internal ID, a token only a human holds, get silently skipped in favor of tools that can run end to end. And tools whose actual responses do not match their declared schema get downranked or dropped. Your API documentation used to be a reference for humans who had already bought. It is now your storefront, your sales call, and your quality signal, evaluated per invocation by a reader with no brand loyalty whatsoever.
Notice what that last property means economically. Hölke's line about the moat evaporating is usually read as a statement about interfaces. The sharper reading is about switching costs: an agent's cost of switching vendors is one routing decision, approximately zero, which means every invocation is a fresh competitive bid. Nothing in the SaaS playbook, not the lock-in, not the learned habit, not the sunk onboarding, survives contact with a buyer that re-evaluates the whole market every time it acts.
Except that today, mostly, it doesn't, and this is the honest part of the audit. The current "market" an agent evaluates is the toolset a human connected: five, ten, thirty tools, chosen upstream by a person or a platform. Cross-vendor comparison, this e-signature contract versus that one, on price, latency, and reliability, happens almost nowhere. The pieces exist in isolation. One peer-to-peer protocol, Hyperspace, ranks discovery results by historical performance, reputation-ranked routing in miniature. Enterprise gateways now log every invocation and enforce access policy through a single governed door, and researchers describe the emerging shape as three tiers: a registry for advertisement, a semantic layer for selection, a per-invocation trust layer for enforcement. But no standard spans the tiers, and the researchers themselves warn that calling it consensus would be premature. There is no agent-native equivalent of a credit rating, no shared ledger of "this endpoint honored its latency contract 99.7 percent of the time last month," no in-band price comparison. The routing table, as Hölke means it, is perhaps one-third built.
So what would the missing two-thirds look like? Follow the four jobs. Evaluation at market scale needs verifiable performance history, uptime, latency, spec-compliance, attested by someone other than the vendor, because agents cannot be charmed but their operators can be defrauded. Transaction needs price signals in-band, which is precisely what the payments layer emerging around HTTP 402 provides: the price arrives in the protocol exchange itself, machine-readable, comparable. (That layer is the next essay in this series.) Feedback needs telemetry to accumulate somewhere that routing decisions can read. And crucially, identity and delegation need to work: the agent must be able to prove who it acts for and what it is authorized to spend, which is what Google's mandate-based authorization work is circling.
Assemble those pieces and you get the next great aggregator, and it is worth saying plainly who is competing to be it. Whoever owns the routing layer owns the demand, the way Google owned the demand of the human web. The candidates are visible: the model providers, whose assistants are the default place agents live; the cloud and edge networks, who see the traffic anyway; the enterprise gateway vendors, who own the corporate policy chokepoint; and possibly an open, federated standard, which is what the registry-plus-protocol crowd is quietly assembling. Hölke's telemetry flywheel predicts the winner: whoever sees the most invocations learns the most about which contracts hold, and routing quality compounds. The routing table is not just infrastructure. It is the position in the stack where margin concentrates once interfaces stop mattering.
If you sell software, you cannot build the routing table, but you can decide how you will rank in it, and the levers follow directly from the mechanics above. Publish a real specification, not a UI with an API bolted on: inputs, outputs, latency envelope, cost per invocation, failure modes. Write tool descriptions the way the selection layer reads them: verb-led, concrete, with declared output shapes you actually honor. Remove every mid-call dependency on a human. Price per operation or per organization, because a routing decision has no line item for seats. Instrument everything, because your invocation telemetry is both your product feedback and, eventually, your rating. And keep your contract true under load, because in a market with zero switching costs, the first failure is the last impression.
I keep this list taped above the desk, metaphorically, because I am building against it: SumoSign's bet is that in e-signatures, the behavior contract, send, sign, audit, verifiably honored, is what an agent will route to, and the incumbent's learned interface is what it will route around. Ask me in a year whether the routing table agreed.
The most quoted prediction in the best essay of the year describes infrastructure nobody has finished building. That is not a weakness in the argument. It is the timestamp on the opportunity. In 1998 the honest answer to "is there search yet?" was "sort of, and the company that answers properly will own the decade." Is there routing yet? Sort of.