Moving the context into a variable does not move the question

A talk from the AI Engineer conference, published September 9, opens with a show of hands and not many go up. Kevin Madura of AlixPartners is there to explain recursive language models, and the short version is this: the model never sees your document. The document sits as a variable in a Python REPL, the model writes code against it, and when a slice is worth reading it hands that slice to another call of itself. The examples are a 200 page invoice, a whole tax code, half a million lines of code, and the closing pitch is that you do not have to worry about context engineering anymore. The AI News issue from the same day carries a related receipt from Harvey and Baseten: on a synthetic diligence benchmark, moving from a plain tool loop to an RLM harness raised the mean rubric pass rate from 23 to 62 percent across models, over data rooms up to 80 million tokens.

Take the pitch at full strength. Context rot is real, stuffing a window is the crude option, and letting the model decide which slice deserves attention beats a chunker somebody tuned once and never touched again.

The vault’s retrieval notes were written before this idea had a name and they still cover it. Term search, vector search, long context, tools, and now a REPL variable are all context construction strategies, not rival religions. What longer context never removed, per that page, is source selection, conflict handling, citations, cost, and evals. A sub-call that returns only what matters is also a sub-call that discarded everything it decided did not, and the same page’s failure list includes the citation-shaped answer that looks grounded while the source says otherwise.

The AI News item adds the quiet part. Harvey’s gains came at least as much from post-training inside the harness as from the harness itself.

So the model stopped needing to see the document. Somebody still has to.