From Agent Loops to Deterministic Graphs: Execution Lineage for Reproducible AI-Native Work

Researchers introduce execution lineage, a DAG-based execution model that makes AI-native workflows reproducible and maintainable by explicitly tracking dependencies and enabling identity-based replay. Tested against traditional loop-based approaches, the system demonstrated superior performance in preserving work integrity during updates while preventing unrelated context contamination.

The research addresses a critical limitation in current large language model deployments: agentic workflows that interleave reasoning and tool use often lack transparent state management, making it difficult to track how changes propagate through multi-step processes. This matters because production AI systems increasingly handle complex, iterative tasks where maintaining consistency across revisions is essential. The execution lineage model represents work as a directed acyclic graph of artifact-producing computations, enabling systems to understand exactly what depends on what and replay changes with precision.

This work emerges from the broader challenge of moving AI systems from one-shot generation toward maintainable, version-controlled workflows. As enterprises deploy agentic systems for document generation, policy work, and complex decision-making, the ability to edit intermediate artifacts without cascading unintended consequences becomes operationally critical. Current loop-based systems can produce high-quality final outputs but mask underlying state inconsistency that compounds across revisions.

The empirical results reveal a fundamental distinction: final-answer quality and maintained-state quality are separate concerns. In controlled policy-memo tasks, DAG replay achieved perfect upstream preservation and downstream propagation while loop baselines frequently contaminated unrelated sections with imported context. This has direct implications for enterprises using AI for collaborative knowledge work, where trustworthiness depends on understanding which changes affect which outputs.

Looking ahead, execution lineage principles could become foundational for AI development platforms and IDE-like tools. The approach suggests a path toward AI-native version control systems that track not just code but semantic artifact lineages. Organizations building internal AI platforms should monitor whether this model influences architectural decisions around composable AI workflows and reproducibility guarantees.

• →Execution lineage uses DAGs to make AI workflows reproducible and change-aware, preserving work integrity across revisions.
• →DAG-based replay prevented unrelated-branch contamination completely while loop baselines frequently imported unwanted context.
• →Final answer quality and maintained-state quality are distinct; high-quality outputs can mask underlying state inconsistency.
• →The model enables perfect upstream preservation and downstream propagation when editing intermediate artifacts.
• →This architecture could influence how enterprises build AI-native version control and collaborative knowledge systems.