Deep agentic search loses to plain semantic retrieval on repo code QA — 46.2% vs 65.2%, at more than double the cost

arXiv·high signalXBlueskyLinkedInCopy link

An empirical study on SWE-QA (arXiv 2608.01507, Aug 2) compares the sub-agent grep pattern now standard in Claude Code, Codex and Antigravity against a pre-built vector index. Semantic search answered 65.2% of repository-level questions correctly versus 46.2% for deep agentic search, and produced each correct answer at less than half the cost. Coding every failed run into a taxonomy showed the sub-agent design did not remove failures but added a new class: 41.8% of its failures — the single largest share — occurred at the planner/sub-agent handoff and were silent, ending in fluent, confident, wrong answers. The context-pollution problem is real, but for read-only questions over an indexable repo the delegation cure is worse than the disease.