July 4, 2026 · Marcus Webb, The Beauty of Mathematics~5 min read
Here's the thing about the million-token context window everybody bragged about all year: you can fill the whole thing, and your coding agent gets dumber. Not slower. Dumber. Somebody finally put it in a newsletter this spring, plain as a flat tire — bigger context windows stopped helping. And the fix, it turns out, is a lesson a search-engine guy could've told you twenty years ago.
More stuff in the box, less work out of it
Claude Opus 4.6 shipped February 5 with a 1-million-token window. A million tokens is a lot of room. The pitch was obvious: dump the whole repo in, hand the agent everything, watch it soar. Except that's not what happened. Stuff the window and the thing loses the plot — misses what's sitting right in the middle, wanders off, hands you code that ignores the one file that mattered. There's a name for it going around now: context rot. More tokens, less signal. The bench got bigger and the work got worse.
I've seen this movie before, in a different shop. You don't help a mechanic by wheeling every tool he's ever owned into the bay and dumping them on the floor. Now he's standing in a pile, kicking wrenches, looking for the one 10-millimeter he actually needs. The tools didn't help. The pile hurt. That's your million-token prompt: a bigger pile is not a better answer.
An old search guy already solved this
Twenty-some years back, Jun Wu wrote a book called The Beauty of Mathematics, mostly about how search engines actually work under the hood. And the spine of the whole thing is one idea that lands like a slap here: more words is not more information. Information is the surprise you didn't already have. Pile on a hundred documents that say nothing new, and you haven't added information — you've added noise. The signal-to-noise ratio drops. The prompt gets fatter and carries less you can actually use.
Search engines figured this out the hard way, decades ago. Nobody wins by handing you all ten million pages that mention your search. The whole trick — the part Wu spends the book on — is ranking. TF-IDF, that clunky little formula: a word that shows up everywhere tells you nothing, a word that's rare tells you what the page is actually about. The engine's real job isn't finding pages. It's throwing almost all of them away and surfacing the three that matter. That's not a side feature. That's the engine.
The teams shipping fast are doing subtraction
So watch what the fast teams in 2026 quietly stopped doing. They stopped running bigger prompts. They built memory stacks that pull only the relevant slice and leave the rest on the shelf. Four layers, roughly: a map of the repo — which symbols call which, what imports what, where the tests are; a record of why the code looks the way it does, the decisions nobody wrote down; a scratchpad the agent keeps for the length of one job; and a shared team memory with permissions on it. None of that is "more context." All of it is better-ranked context.
And the numbers back the plain-English version. One 2026 memory benchmark clocked a token-efficient setup at 92.5 on the LoCoMo test while pulling only about 6,956 tokens per call. Read that twice. High accuracy, on a sliver of the context. It didn't win by seeing more. It won by seeing the right part. That's the whole game, and it's exactly the game the search engines have been playing since before any of us had a coding agent.
What this means for you on Monday
If you've been jamming your whole codebase into the prompt because the window finally lets you — stop. That's the pile-of-tools move, and it's costing you accuracy you can't see. The better instinct is the search engine's instinct: what are the three files this task actually touches? Feed those. Leave the rest on the shelf.
Turns out the skill that separates a working setup from a bloated one isn't cramming. It's cutting. Knowing what to leave out is the job — same as it was for the search guys, same as it's always been. A great model, like a great search engine, isn't a great memory. It's a great filter. The pipe that flows is the one that isn't clogged.
The win isn't a bigger window; it's a better filter
Stuffing more context in lowers the signal — the teams pulling ahead surface only the relevant slice, which is the oldest trick in the search-engine book.
Here's the plain version of it. For a year the pitch was capacity — look how much fits now. Capacity was never the constraint. Relevance was. The million-token window is a bigger warehouse, and a bigger warehouse full of junk is still full of junk. The move that pays is the boring one Jun Wu wrote down decades ago: rank hard, keep the three that matter, throw the rest back. Do the subtraction. That's not a smaller job. That's the whole one.
Framing drawn from Jun Wu, The Beauty of Mathematics. Benchmark and product figures reflect 2026 reporting (a coding-agent newsletter and a memory benchmark report) — treat them as the read on a fast-moving field, not settled engineering law. This is one working engineer's take, not a verdict on any specific tool.
Brute force fills the million-token window and the signal ratio collapses ("context rot" / lost in the middle); retrieval ranks the context and surfaces only the relevant slice — one 2026 memory benchmark hit LoCoMo 92.5 at ~6,956 tokens per call. Framework: Jun Wu, The Beauty of Mathematics. One reading of a fast-moving field, not settled law.