LongCapHarness boosts long-context reasoning without training

This is less a model breakthrough than a strong case for treating reasoning as software infrastructure around the model.

• –Explicit primitives such as lookup, ordering, state tracking, and relation following handle operations models often struggle to maintain across long contexts.
• –Gains are strongest on recall and RULER, where code-controlled execution can replace fragile in-context bookkeeping.
• –Semantic tasks remain harder: LongCapHarness beats PolicyLong on LongBench-v2’s Medium subset but trails it on the Long subset.
• –Development costs are estimated below half those of competing post-training approaches.
• –The “no training” claim has an important caveat: GPT-5.6-Sol constructs the harness, and weaker models struggle to build equally effective workflows themselves.