- Who it helps
- Researchers and developers measuring retries, corrective prompts or agent recovery logic.
- Availability
- The September 30 paper is publicly readable. An executable reproduction was not verified during this sweep.
Why the starting state matters
An agent can succeed after a retry without the retry being responsible. The study uses paired continuations to separate rescued runs from cases where an intervention harms a run that would otherwise succeed.
When Harnesses Lose the Signal: Causal Evaluation of Recovery in LLM Agents ↗
A narrow reported result
The authors study Qwen3-14B in ALFWorld and report success increasing from 70.33 to 73.33 percent. That is three percentage points in this setup; it does not establish a general benefit for every agent or task.
When Harnesses Lose the Signal: Causal Evaluation of Recovery in LLM Agents ↗
What you can do
- Before adding recovery logic to a workflow, save a failing state and define how to compare its subsequent outcomes fairly.
Where the evidence stops
The reported effect is specific to the model, environment and evaluation protocol. No experiment was reproduced here.
Sources read. Sources checked October 18, 2026. No product workflow or benchmark was executed.
Sources & further reading
Paper · paper
When Harnesses Lose the Signal: Causal Evaluation of Recovery in LLM Agents ↗
Selected for direct evidence about the subject, its availability or the limits of the claim.
Paper authors · page checked 2026-10-18
Coverage note: Independent hands-on comparisons and current cross-platform discussion are incomplete in this bounded sample. X public search required login.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
This paper tests recovery interventions by comparing what happens from the same agent execution state with and without an intervention.