The authors introduce the concept of an agentic transaction to address challenges in autonomous agent execution, such as reliable outcomes and safe concurrency. They propose a framework that reinterprets classical ACID properties into four semantic guarantees: Semantic Atomicity, Consistency, Isolation, and Durability.

  • The system realizes these guarantees through transactional exploration-execution-validation cycles and transactional skill hubs.
  • It employs confidence divergence-based validation to ensure accuracy during execution.
  • Semantic dependency-aware isolation prevents interference between concurrent agent operations.
  • Transaction-aware semantic state management ensures durable state updates.

Experimental results on widely used benchmarks show that the system achieves a 10.6% improvement over state-of-the-art agents, including Claude Code.