Can externalized bookkeeping let smaller search agents beat larger ones?

The usual framing of a search agent is a policy over a growing transcript: the model must simultaneously decide what to search and remember what it has seen, which evidence is useful, which constraints remain open, and which claims it actually checked. Harness-1 argues this overloads reinforcement learning — it forces the policy to optimize both genuine semantic search decisions and routine bookkeeping that the environment can maintain far more reliably.

The fix is a division of labor. The harness maintains environment-side working memory: a candidate pool, an importance-tagged curated set, compact evidence links, verification records, deduplicated observations, and budget-aware context rendering. The policy keeps only the semantic decisions — what to query, what to keep or discard, what to verify, and when to stop. A 20B model trained this way reaches 0.730 average curated recall across eight benchmarks, beating the next open searcher by +11.4 points and staying competitive with much larger frontier models.

The deeper claim is that the harness is not an implementation detail but part of what the policy learns to use — gains transfer to held-out benchmarks and survive component ablation. This is the search-agent instantiation of a broader principle: capability moves out of parameters and into the editable scaffolding. Since Is long-context bottleneck really about memory or compute?, externalizing bookkeeping is exactly what frees the policy's scarce reasoning compute for decisions only it can make.