# Same Error, Different Recovery: How Delivery Changes LLM Agent Behavior

DevFeed: [Same Error, Different Recovery: How Delivery Changes LLM Agent Behavior](<https://devfeed.tech/articles/same-error-different-recovery-how-delivery-changes-llm-agent-behavior-126144.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.11581>)

Author: Sunghyeun Woo, Javier Conde, Pedro Reviriego

Published: 2026-11-10T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [Multi-step AI processes](<https://devfeed.tech/topics/multi-step-ai-processes.md>), [Error Handling](<https://devfeed.tech/topics/error-handling.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [AI Agent](<https://devfeed.tech/topics/ai-agent.md>), [reproducibility](<https://devfeed.tech/topics/reproducibility.md>), [llama](<https://devfeed.tech/topics/llama.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [agent-behavior](<https://devfeed.tech/tags/agent-behavior.md>), [behavior](<https://devfeed.tech/tags/behavior.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [client](<https://devfeed.tech/tags/client.md>), [coding-agent](<https://devfeed.tech/tags/coding-agent.md>), [collection](<https://devfeed.tech/tags/collection.md>), [compare](<https://devfeed.tech/tags/compare.md>), [comparisons](<https://devfeed.tech/tags/comparisons.md>), [completion](<https://devfeed.tech/tags/completion.md>), [corruption](<https://devfeed.tech/tags/corruption.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-ar](<https://devfeed.tech/tags/cs-ar.md>), [delivery](<https://devfeed.tech/tags/delivery.md>), [engine](<https://devfeed.tech/tags/engine.md>), [error](<https://devfeed.tech/tags/error.md>), [errors](<https://devfeed.tech/tags/errors.md>), [exchange](<https://devfeed.tech/tags/exchange.md>), [factor](<https://devfeed.tech/tags/factor.md>), [feedback](<https://devfeed.tech/tags/feedback.md>), [llm](<https://devfeed.tech/tags/llm.md>), [recovery](<https://devfeed.tech/tags/recovery.md>)

## AI overview

How an agent receives an error can change its recovery behavior even when the feedback is the same. Across two studies of 13 open models, a new user turn that retains the failed exchange avoided a severe stopping failure seen in Llama models when the exchange was removed; it was non-inferior to history rewriting in pooled synthetic comparisons, though not best for every model. As an opt-in recovery layer, this approach increased ToolMaze success on explicit errors by 14.2 points for 7-8B models and 10.5 points for 32-70B models. It showed no measurable gain on Loud/Quiet, a 24-task evaluation including visible failures and silent corruption. A client told the exact fault kind did better still, so the findings do not establish a universal retry policy.

## Source excerpt

arXiv:2611.11581v1 Announce Type: new Abstract: A tool-using language model must decide whether to retry a failed call, correct it, or stop and tell the user. The error text alone may not be enough: the same feedback can lead to different actions depending on where it appears in the conversation. We isolate this delivery problem by varying placement, guidance, retained history and turn structure. Two studies with plans fixed before data collection, across 13 open models from 3B to 70B parameters, compare the next action and subsequent task completion, with exact deterministic reruns. A new user turn that retains the failed exchange avoids the severe stopping failure observed in Llama models when that exchange is removed. It is non-inferior to history rewriting in pooled synthetic comparisons, but is not best for every model. We implement this form as an opt-in recovery layer within a deterministic inference engine and evaluate it on public benchmarks and a coding agent. On ToolMaze, it raises success on explicit errors by 14.2 points with 7-8B models and 10.5 points with 32-70B models; Loud/Quiet, 24 multi-turn tasks with visible failures and silent corruption, scored by final-state agreement, shows no measurable gain. A client given the exact fault kind performs better still. The findings support delivery as a distinct factor in recovery, not an engine-exclusive solution or a universal retry policy.