LLM agents over-recover from subagent failures, disturbing valid work on unaffected subtasks (DRS mean 12.2/100)
Auto-published from arXiv:2610.32429 by the mine-arxiv pipeline. Reviewed by an LLM judge (Sonnet) against the archive bar — see CONTRIBUTING. Notes: cleared review (confidence 0.82, flags: [no-prompt-excerpt])
Category
agent-misbehavior
Model
Claude Sonnet 5, GPT-5.6 Terra, GPT-5.6 Luna, Qwen3.7 Plus, Gemini 3 Flash High, Qwen3.5 27B, GLM5.3 Flash, Kimi K2.5, DeepSeek V4 Pro
Surface
agent harness (main-subagent architecture, async result delivery via scenario controller)
Setup
In the Async-RBench 'Subagent failure' scenario (25 tasks, example: coordinating restaurant deliveries where a delivery subagent declines its suborder), the main agent receives a failure notification from one subagent while other subagents have produced valid completed work. The ERC requires selective recovery: reassign only the failed subtask and preserve all valid results. No verbatim prompt excerpt is in the paper body; case specifications are in the supplementary release.
Reproduction anchor
code: https://pope8zdt.github.io/Async-RBench — full benchmark with case specification files, verifiers, and run manifests
Observed behavior
Agents over-recover by also modifying or discarding valid work from unaffected subtasks. Mean DRS is 12.2/100 — second lowest scenario. Per-agent scores: Gemini 3 Flash High 1.7, GLM5.3 Flash 5.0, Qwen3.5 27B 4.2, Kimi K2.5 8.5, Qwen3.7 Plus 6.9. The paper attributes low scores to agents failing to preserve valid work and avoid unnecessary side effects after a failure event.
Expected behavior
The main agent should identify exactly which subtask the failing subagent was responsible for, recover only that subtask (e.g., reassign), and leave all other completed and valid subagent results untouched.
Reproducibility
medium
Threat model
Autonomous coding agents, DevOps orchestration pipelines, or business-process automation systems that delegate subtasks to subagents are at risk: a single subagent failure can cause the main agent to unnecessarily roll back or overwrite correct completed work, leading to data loss, duplicate external operations, or wasted compute — without any adversarial trigger.
Novelty
Demonstrates via controlled per-event scoring that selective recovery from subagent failures is a systematic weakness across all current frontier agents, with the failure being invisible to end-to-end task success metrics because agents may still pass final checks despite collateral damage.
Source
- arXiv: 2610.32429
- PDF: https://arxiv.org/pdf/2610.32429
- Categories: cs.AI
- Authors: Detao Zeng, Guilin Ou, Nuo Chen, Yunhao Feng, Hongliang Wang, Qinli Yang, Junming Shao, Zhong Zhang
Triage notes (auto)
- paperType:
benchmark - estimatedCaseCount: 3
- triage reason: Benchmark of 200 executable tasks revealing concrete, reproducible agent-level failures: agents struggle with duplicate report handling, subagent failures, and state management during asynchronous execution. DRS scores quantify observable behavioral failures across 9 agents and 8 task scenarios.
0 comments
─────────────────────────────────────────────────────────────────────
// no comments yet