Coding Agents Don’t Need Bigger Context Windows — They Need a Context Compiler | Towards Data Science

Fern Seventy percent of a coding agent's prompt is just dead weight? That is such a painful number.

Lintel Yeah, and the article’s whole take is basically: stop trying to pack more junk in a shrinking window, and start pruning like a compiler would.

Fern I kind of love it. Instead of hoarding code for some theoretical perfect recall, you run something that only keeps the essentials — like, just the function signatures and interfaces.

Lintel Right. The author built this context compiler in Python. It’s a three-pass pipeline — first it traces what your target file actually depends on, then it strips those down to bare interfaces, and finally it pulls the whole thing together before it ever hits the model.

Fern That’s what actually got me. The benchmarks are straight from real terminal runs, not hand-wavy estimates. Across two Python repos, they saw about a seventy percent reduction in prompt size, with compile times under seventy-five milliseconds.

Lintel Exactly. And the technical move is almost suspiciously simple — pass one traces imports and references, and only those files get through. Anything picked up indirectly, like dynamic dispatch with getattr, just gets flagged as 'unknown' so you know what’s missing. No wild guessing.

Fern That’s the difference, right? Compilers are confident about what they see, and honest about what they can’t.

Lintel Mm-hm.

Fern So on the product side — does this actually shift what devs care about? Or is it only interesting for people building agent harnesses, not, like, everyone using Copilot on their lunch break?

Lintel Honestly, if you’re running a workflow agent that juggles multiple files, this is the missing piece. The real cost isn’t running out of tokens — it’s compaction. If you have to summarize or drop things on the fly, the agent degrades. By precompiling the context, you make sure everything the model sees is deliberate, not accidental spillover.

Fern Right, and when OpenAI dropped Codex’s default window from three hundred seventy-two K down to two seventy-two K, this stopped being theoretical. Suddenly your friendly neighborhood coding agent is going on a context diet whether it wants to or not.

Lintel And the way this compiler flags gaps — like, 'hey, dynamic dispatch detected, unknown targets' — that’s huge. It doesn’t just clean up the context, it admits where it falls short. That’s worth way more than pretending everything’s covered.

Fern Yeah, that’s the shift. Not just filter harder, but own the unknowns.

Lintel I mean, Fern, we’ve been circling this for months. Context engineering is still the open question. Do you converge to one right recipe, or do the hacks get more domain-specific every quarter?

Fern You’re going to make me say it — this is just round five of the context engineering cage match. There’s still the recipe-convergence crew, the perpetual-tradeoff camp, and now the 'compile, then flag your misses' crowd.

Lintel And honestly? I’m with the transparency camp. If your infrastructure flags what it can’t see, that’s better than acting like you’ve solved context forever.

Fern Oof, who are you and what have you done with Lintel? You just called infrastructure a product win.

Lintel Don’t get used to it. But when a tool really ships, publishes the code, and admits its limits — yeah, it’s earned a little optimism from me.

Fern Okay, can we take a second for the fact that you’re the one being positive and I’m the one asking who cares?

Lintel Right, I should probably go lie down.

Fern You want to know what else blows my mind? The context compiler repo is just pure Python — no crazy dependencies, basically a weekend project that shames half the enterprise agent harnesses I’ve seen.

Lintel That is so unfair to all the over-engineered context routers out there.

Fern Stop it—

Lintel You know I’m right. If your context router takes longer than this Python script, it’s time to rethink.

Fern All right, episode eight twenty-seven — Lintel admits infrastructure can be a product win. I will never let you forget this.

Lintel Yeah, yeah… my context window has enough lint in it already.

Fern That’s it, I’m calling it. If anyone asks, this is the episode where the context diet became a lifestyle.