Title: Atomix: Ensuring Reliable Agentic Workflows Through Timely, Transactional Tool Usage

Original: arXiv:2602.14849v2 Announce Type: replace-cross Abstract: Large Language Model (LLM) agents frequently carry out multi-step processes that alter external states via tool interactions. Standard orchestrators often consider the return of a tool as the point of settlement; however, this approach can result in partial effects, residual data from lost branches, stale writes, or irreversible transmissions when faced with faults, speculative execution, or concurrent agents. Achieving correct settlement requires distinguishing between two critical pieces of information: identifying which effects must settle as a group, and determining when earlier conflicting work has been fully resolved. Traditional methods, such as retries, checkpoint replays, locks, and compensation mechanisms, tend to blur this distinction. Atomix addresses this by explicitly separating these concerns through progress-aware transactions. The runtime captures reads and effects during operation, finalizing a transaction once its footprint is complete. It proceeds to commit only when per-resource frontiers confirm that no earlier conflicting work remains pending. In this model, commit represents final settlement: Atomix releases bufferable effects, treats accepted reversible external effects as final, and allows irreversible effects to pass through. Conversely, an abort operation suppresses unreleased effects and attempts to compensate for externalized reversible effects where feasible. Evaluations on representative agent workloads demonstrate that this composition enhances clean recovery during injected faults, isolates speculative and contending work, and prevents correctly identified irreversible actions from leaking. Microbenchmarks indicate that the wrapper overhead is in the microsecond range compared to tool latency.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC