Search papers, labs, and topics across Lattice.
The paper introduces Atomix, a runtime system that provides transactional semantics for LLM agent tool calls to address the problem of unintended side effects from speculative or failed agentic workflows. Atomix uses epoch-based tagging of tool calls, resource frontier tracking, and progress predicates to determine when to commit actions, enabling delayed buffering or compensation of externalized effects. Experiments on real-world workloads with fault injection demonstrate that Atomix improves task success through transactional retry and enhances isolation under speculative execution and contention via frontier-gated commits.
Prevent unintended side effects from LLM agents with Atomix, a runtime that brings transactional guarantees to tool use.
LLM agents increasingly act on external systems, yet tool effects are immediate. Under failures, speculation, or contention, losing branches can leak unintended side effects with no safe rollback. We introduce Atomix, a runtime that provides progress-aware transactional semantics for agent tool calls. Atomix tags each call with an epoch, tracks per-resource frontiers, and commits only when progress predicates indicate safety; bufferable effects can be delayed, while externalized effects are tracked and compensated on abort. Across real workloads with fault injection, transactional retry improves task success, while frontier-gated commit strengthens isolation under speculation and contention.
