Skip to content

Silent Handoff Failure in Delegated Code Search

Delegating repository search to a subagent produces a confident wrong answer when the handoff breaks, and the breakdown is silent 91% of the time.

A silent handoff failure is a delegated search that returns nothing useful while the system reports success. Across 720 repository code questions, breakdowns at the planner-to-subagent handoff accounted for 41.8% of the delegated architecture's failures and 0% of the retrieval architecture's. In 91% of those cases "the agent still returned a fluent, confidently worded answer rather than reporting that anything had gone wrong, so the breakdown was not apparent from the output alone" (Rafiei Oskooei et al., 2026).

Where the finding applies

The study fixed three conditions, and none of them describes a general coding session (Rafiei Oskooei et al., 2026):

  • The task is read-only question answering. Nothing here covers patch generation, test execution, or refactoring.
  • The repository can be indexed ahead of time. The 15 SWE-QA repositories were frozen public Python projects, so the index was never stale.
  • Each question runs in a fresh session, so the design cannot score the benefit delegation exists for: an orchestrator context that stays clean across many tasks.

What the comparison measured

Two architectures answered the same 720 questions. One gave a ReAct agent vector retrieval over a pre-built index; the other gave a planning orchestrator a subagent that explored the repository in an isolated context and returned a condensed result (Rafiei Oskooei et al., 2026).

Model Semantic index Delegated search Gap
Gemini 2.5 Flash 48.4% 42.8% 5.6 pp
Gemini 2.5 Pro 54.2% 44.2% 10.0 pp
Qwen3-235B 68.8% 58.2% 10.6 pp
Gemini 3 Flash 89.3% 39.7% 49.6 pp
Pooled 65.2% 46.2% 19.0 pp

Retrieval also cost $0.32 per correct answer against $0.74 for delegation, and delegation matched or beat it on 1 of the 15 repositories (Rafiei Oskooei et al., 2026). Plan against your own model, not the pooled row: the same choice buys 5.6 points on one model and 49.6 on another.

Delegation still located code well. Localization misses were 24.4% of its failures against 53.6% of retrieval's (Rafiei Oskooei et al., 2026). Applying each share to that architecture's own failure rate puts the absolute miss rate near 13% of questions for delegation and 19% for retrieval, so the accuracy gap traces to the handoff.

Why it works

Delegation compresses twice, and the second compression destroys the evidence needed to audit the first. The planner sends a natural-language brief down to a subagent working in an isolated context, and the subagent sends a condensed result back up. Everything in between goes away with the subagent's context window: the queries it ran, the files it opened, the candidates it rejected. The planner cannot separate a well-scoped brief answered correctly from a mis-scoped brief answered plausibly, so it writes an answer from the summary either way (Rafiei Oskooei et al., 2026).

Vector retrieval has no equivalent seam. It returns code blocks into the answering agent's own window, so a miss arrives as thin or irrelevant evidence that the same agent can weigh. Retrieval failures therefore concentrate in localization, which is visible, rather than in coordination, which is not.

When this backfires

An index-first default is wrong in four situations.

  • The tree is volatile or unindexable. The study names filesystems "too large, too volatile, or too arbitrary to index ahead of time" as the case delegation is built for (Rafiei Oskooei et al., 2026). Staleness costs most when the agent reads its own writes: it searches for text it just added, misses, and "it'll often go into a wild goose chase, waste tokens" (Cursor).
  • The session is long and spans many tasks. Isolation keeps the orchestrator's window clean over hours of work, which a single-question benchmark cannot register (Rafiei Oskooei et al., 2026).
  • The work writes or executes code. The benchmark is read-only, so none of the numbers transfer.
  • The task is genuinely parallel exploration. Anthropic measured a lead agent with subagents beating a single agent by 90.2% on open-ended research, while noting that "most coding tasks involve fewer truly parallelizable tasks than research" (Anthropic).

The study's conclusion is a router: keep "a cheap semantic index for retrieval" and reserve delegated exploration "for the questions, repositories, or tasks that genuinely require it" (Rafiei Oskooei et al., 2026).

Key Takeaways

  • Treat a search subagent's answer as unverified until the citations it returned are checked against the files, because a broken handoff reads exactly like a good one.
  • Require file paths and line ranges back from a search subagent rather than prose, so the planner holds something it can re-open.
  • Size the architecture decision against your own model. The measured gap ranged from 5.6 to 49.6 percentage points across four models on identical questions.
  • Delegation localized code more reliably than retrieval on this benchmark. Its accuracy loss came from the handoff contract, which is the part you control.
  • Non-terminating loops took 13.5% of delegated failures against 2.9% of retrieval's, so budget a turn ceiling on any exploration subagent (Rafiei Oskooei et al., 2026).
Emerging — single recent source Last reviewed: 2026-08-06