# Interface Surface Area Costs Selection Accuracy, Not Just Tokens ## Claim Expanding the set of options presented to a decision-maker degrades the quality of the choice, so surface area must be curated rather than merely paid for. <!-- atomicity: 1 idea ✓ --> ## Evidence **Source**: [[3. Resources/Readwise/Articles/The New MCP Roadmap|The New MCP Roadmap]] — MCP Core Maintainers, stating the justification for a new priority area > [!quote] > Connecting to a server with a hundred tools means the model pays for that entire surface before the user has asked a single question, and tool selection tends to get worse as the list grows. Two costs are named in one sentence and they are different in kind. The first is the token toll — already well understood, and the subject of [[3. Resources/Permanent Notes/A Declared Capability Costs You Every Turn Not Every Use]]. The second is **correctness**: the model's ability to pick the right tool degrades as the catalogue grows, independent of what the catalogue costs to carry. That second cost is the more damaging one because it is invisible. A token bill announces itself. A selection that quietly gets worse looks like the model getting dumber, or the integration being flaky, or the prompt needing work — anything except the number of tools on offer. The protocol's response is **progressive discovery**: a server offers a small entry point and reveals more of its catalogue as the conversation narrows. Notably, clients were already approximating this by deferring tool schemas and loading them on demand, so the specification is codifying practice rather than leading it. ## Scope & conditions - Holds for models selecting among tools and, by the same mechanism, for humans selecting among options — this is choice overload with a machine in the chooser's seat. - Weakens where options are strongly typed or hierarchically organised so the *effective* choice set stays small at any moment. That is precisely what progressive discovery attempts, which is why it is the right shape of fix. - **Currently asserted, not measured.** The MCP roadmap offers no evaluation data for the accuracy claim, and it is load-bearing for a whole priority area. Hold as a working hypothesis: the clean test is the same task set with the catalogue padded by irrelevant tools, scoring selection against ground truth. - The practical consequence is a design rule rather than a budget line: curate what you expose, don't just price it. ## Links - [[AI Agent Safety MOC]] — a scaling limit on declared capability that degrades tool choice, not just cost - [[AI Cost & Tokenomics MOC]] — the correctness cost sitting beside the per-turn token toll, and the one that stays invisible - [[3. Resources/Permanent Notes/A Declared Capability Costs You Every Turn Not Every Use]] — the token half of the same problem; that note prices the surface, this one says the surface also damages the decision - [[3. Resources/Permanent Notes/Structured Tool Access Bounds AI Attack Surface on Production Systems]] — a third cost of the same surface, security rather than accuracy or tokens - [[3. Resources/Permanent Notes/Narrow Agent Prompts Beat Broad Exploratory Ones Once the Task Is Named]] — the prompt-side twin: narrowing the ask improves output for the same reason narrowing the catalogue does `proof_state: working` - [[WordPress Platform Design MOC]] — the scaling limit a growing ability registry runs into - [[3. Resources/Permanent Notes/Agent Memory Is Written Far More Often Than It Is Read]] — because applies-in: a memory catalogue degrades retrieval the way a tool catalogue degrades selection, and the write/read ratio is the instrument that makes the growth visible. - [[3. Resources/Permanent Notes/Instruction Files Are the Last Rung Not the First]] — because applies-in: the ordering rule that keeps surface area from growing in the first place, by pushing rules down to a rung the model never has to select among. - [[3. Resources/Permanent Notes/A Release Lead's Output Is Mostly Subtraction]] — because applies-in: every feature not added is surface area not spent, which is the cost this note prices. - [[3. Resources/Permanent Notes/Exfiltration Needs Three Capabilities Together Not Any One of Them]] — because extends: the security cost of enabling more surface than the task needs, and the one that does not degrade gracefully.