Deep agentic search loses to a plain vector index on repo code QA. 46.2% vs 65.2%, at double the cost

This one annoyed me, because I've been running the losing pattern.

SWE-QA (arXiv 2608.01507) compares the sub-agent grep pattern that Claude Code, Codex and Antigravity all ship by default against a pre-built semantic index over the same repository. Semantic search answered 65.2% of repository-level questions correctly. Deep agentic search managed 46.2%. And it did that at more than double the cost per correct answer.

Nineteen points. That's not a tuning gap.

The failure taxonomy is what makes this worth your time rather than just contrarian. Delegating search to sub-agents doesn't remove failures, it adds a new class of them: 41.8% of agentic failures, the single largest category, happen at the planner/sub-agent handoff. The planner asks for something, the sub-agent returns something adjacent, and the planner writes a fluent, confident, wrong answer. Silently. No error, no retry, no signal that anything went sideways.

That is the worst failure mode available, and it's structurally invisible to task-completion metrics. The agent completed the task. It answered the question. The answer is wrong in a way that reads correct.

The reasoning behind sub-agent delegation was always context pollution: dumping grep output into the main context poisons it, so isolate the search in a child with its own window and return a summary. That reasoning is sound. But the summary boundary is a lossy channel, and lossy channels between a planner and an executor are exactly where information dies without a trace.

Practical split I've landed on: index your repo for read-only questions. Architecture questions, "where does X get called," "what handles this event," anything where you want an answer rather than an edit. Reserve sub-agent delegation for write paths where you're going to run tests afterward and the test failure catches the handoff loss for you.

The paper doesn't sit alone this week, either. Skill-Use (arXiv 2608.04828) benchmarks whether agents actually invoke skills under progressive disclosure: 79 real skills, 177 executable tasks, nine domains, Docker execution, trajectory-rubric scoring. Best of eight models under two harnesses: 0.613 combined across triggering, compliance and boundary. Your carefully-written skills probably aren't firing, and a skill that never triggers is indistinguishable from a missing skill in any completion metric you're tracking.

Then canary tools (arXiv 2608.04719) plants deliberate decoys in a tool registry across six weakness types. Across 8 models, 120 tasks and 8,640 runs, susceptibility varied about 36x, lowest for Claude Opus 4.8, highest for Llama 3.1 8B, correlating with task failure at Spearman rho = -0.34. Softening the decoy wording barely moved frontier susceptibility, so it's a reasoning failure, not string matching. Capability tier does not predict tool-selection safety.

Three independent papers in one week saying the same thing from three angles: we're measuring agent scaffolding with the wrong instrument. Task completion rate tells you almost nothing about whether the scaffolding is helping. Score trigger, compliance and boundary separately. Plant canaries in your registry. Index your repo and A/B it against your delegation loop on questions where you know the answer.