Fault-injection study separates idempotency from recovery in language-model agent tool actions
An arXiv paper reports that retrying interrupted agent tool calls risks duplicate effects, while assuming success can hide actions that never happened, and measures the tradeoffs.
In plain words
When an AI agent's tool call is interrupted, the target system may already have changed, so retrying can duplicate the action and skipping the retry can leave it undone. The authors built a test runtime with safeguards like an action ledger and reconciliation, then injected faults to see which designs actually recover. Their results show duplicate suppression and recovery are distinct properties with measurable tradeoffs.
What was announced
The study, published as an arXiv abstract, describes a controlled agent runtime combining authorization, an action ledger, target-side idempotency, reconciliation, and evidence sealing. The authors injected ten single-host fault classes and compared four configurations and three ablations, judging outcomes with a target-side oracle kept separate from the runtime ledger. Each Windows and Linux run covered 7,513 trials built from 96 captured model responses across 24 constructed cases.
Across 200 known crash-and-replay pairs, the full configuration recorded no repeated effects, while each simpler baseline produced 117. Removing a ledger-related component left 170 orphaned effects, and disabling reconciliation left 100 post-effect crash outcomes ambiguous even though no duplicates were observed. The full runtime also ran at roughly one-seventh of the fail-open configuration's throughput, and injected evidence-write faults left 75 of 100 trials per platform unresolved. A supplement designed after the initial results, reusing the same inputs, found a similar pattern under response loss or delay. The authors present these as measured reliability and performance tradeoffs within the tested benchmark only.
What it does not show
- The evidence is the paper's abstract only; full methodology, threat model, and peer review status are not available in the supplied passages.
- All results come from a single-host benchmark with constructed cases and captured model responses, so they may not generalize to production agent deployments.
- The abstract reports aggregate counts without confidence intervals or statistical tests, and no independent replication is cited.
For security and IT teams
- Scope
- Findings apply to the tested single-host benchmark with ten injected fault classes on Windows and Linux, not to arbitrary agent deployments.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.