# How an agent chooses between two services that do the same thing An agent comparing equivalent services filters on what it can actually pay for, then ranks on price and latency, then on whatever evidence of reliability is available. Brand and familiarity carry no weight. A seller wins by being legible in the comparison, which means a fixed price, a readable interface and a public history. Published: 2026-10-14 Last reviewed: 2026-10-14 Publisher: Roundhouse AI Ltd. Source: https://roundhouseai.io/resources/how-agents-choose-between-services --- ## What order does the comparison happen in? An agent with a task and two candidate endpoints does three things, and most candidates are eliminated in the first. **1. Filter on feasibility.** Can I pay on this network? Do I hold this asset? Is the amount inside my budget? Do I implement this scheme? Can I parse the interface? Anything failing these is not a candidate at any price. **2. Rank on cost and speed.** Price, which is directly comparable when it is a fixed number and not when it is a ceiling. Latency, which includes settlement time if the seller settles before serving. **3. Break ties on evidence.** Whatever the agent can establish about whether this seller actually works: a public payment history, a registry identity, a prior successful interaction of its own. | Stage | What loses you the sale | | --- | --- | | Feasibility | Wrong chain, unparseable interface, no free discovery path | | Ranking | A higher price, or a ceiling where a competitor shows a number | | Tie-break | No history, thin counterparty breadth, self-payments | ## Why do most sellers lose at stage 1? Not on price. On being impossible to evaluate. A buyer that cannot read your schema without paying cannot tell whether you do what it needs, so it moves on. A buyer whose funds are on a chain you do not accept cannot pay you at any price. A buyer that gets a bare 402 with no terms has nothing to act on. These are all cheap to fix and all commonly broken: - **Keep a free discovery path.** A schema endpoint, a health check, a short description. Unpaid. - **Return full terms in your 402.** Scheme, network, asset, amount, payee, resource, timeout. - **Accept the chain your buyers are on**, and list more than one entry if you can. - **Publish an OpenAPI document** at a predictable path, so "can I use this" is answerable without prose. ## Stage 2 rewards legibility as much as level A fixed price can be compared instantly. A ceiling cannot. Two sellers advertising `exact` at two thousand and three thousand smallest units are ranked in one comparison. A seller advertising `upto` fifty thousand looks more expensive than both to anything comparing naively, even if it typically settles at a thousand. [Exact against upto](https://roundhouseai.io/resources/x402-exact-vs-upto-schemes) covers when the ceiling is unavoidable and what to publish alongside it. The same applies to tiers. A tier table assumes a relationship and a human to read it. An anonymous buyer sees one number or gives up. And get the units right. An amount in smallest units published as dollars, or the reverse, makes you look a million times cheaper or more expensive than you are. [Pricing for agents](https://roundhouseai.io/resources/how-to-price-an-endpoint-for-agents) covers the trap, which is the most consequential pricing error in the ecosystem because it fails silently. ## Stage 3 is the one you cannot fake cheaply When two endpoints are feasible and similarly priced, the tie-break is evidence, and the evidence available is your public settlement history. What a careful buyer reads: how many distinct counterparties have paid you, over how long, in what distribution of amounts, with what rate of self-payment. [Auditing a service's traffic](https://roundhouseai.io/resources/how-to-audit-an-x402-services-traffic) is the buyer's side of exactly this. Breadth is what counts. Volume is cheap to manufacture and counterparty breadth is not, so an automated screen that weights volume is a screen worth gaming and one that weights distinct payers is not. Expect buyers to converge on the latter. > [!IMPORTANT] > This inverts a familiar instinct. In a human market, a big headline number is the thing to publicise. In a market where buyers audit the chain, a big number from four addresses is worse than a small number from four hundred, because the first one is visibly manufactured and the second is visibly not. ## What has no effect at all? Worth saying explicitly, because effort spent here is wasted. - **Brand.** Not represented in the comparison at all. - **Design.** The buyer is not looking at your page. - **Marketing copy.** Unless it is in a catalogue description an agent reads, which is one or two sentences. - **Switching cost.** There is none. A buyer that used you yesterday re-evaluates today if it is designed to. The last one cuts both ways and the upside is larger than it looks. You can win a buyer instantly, with no sales cycle, by being the better option at the moment of comparison. The same absence of inertia that loses you incumbency gives you entry. ## Frequently asked questions ### Should I always be the cheapest? No. Be feasible, legible and clearly described, and price for a sustainable margin. Agents that pick purely on price also pick purely on price when something cheaper and worse appears, which is not a customer base worth buying. ### How do I signal quality if the buyer cannot see my output? A free sample call is the strongest signal and costs you little. A public payment history is the second. Claims in a description are the weakest and are usually ignored. ### Do agents remember a bad experience? If the operator built memory, yes, usually as a local preference rather than anything public. There is no shared reputation layer for delivery quality, and the registry-based attempts at one have not demonstrated it works. [ERC-8004](https://roundhouseai.io/resources/erc-8004-explained) covers that research. ### Should I compete on latency? It matters more than most sellers expect, because an agent's own response time includes yours. If you settle before serving, your settlement time is part of your latency and worth measuring at the tail rather than the median. ### What if my service is genuinely better but more expensive? Say what the difference is, in your catalogue description and your schema, in terms a program can act on: coverage, freshness, accuracy. A quality difference the buyer cannot represent is a quality difference it will not pay for. ## Where to check this yourself [/services](https://roundhouseai.io/services) is the comparison a buyer sees, with listed prices and observed activity side by side across indexed endpoints. [/transactions](https://roundhouseai.io/transactions) shows settled amounts rather than advertised ones, which is where you find out what buyers are really paying in your category. --- Roundhouse indexes x402 settlements and ERC-8004 identities into one public record. - Every article: https://roundhouseai.io/resources.md - Every surface on this site: https://roundhouseai.io/llms.txt - The index, endpoints and schema: https://api.roundhouseai.io/v0 - A key of your own, $0.01 over x402, no account: https://api.roundhouseai.io/v0/test/x402 - Limits and the SQL surface: https://roundhouseai.io/docs/api