From Agent Loops to Deterministic Graphs: Execution Lineage for Reproducible AI-Native Work
Researchers Josh and Seth Rosen introduce 'execution lineage,' a new execution model designed to enhance the reproducibility and maintainability of AI-native workflows. Current large language model systems often rely on implicit conversational states within agentic loops, leading to instability when preserving work products or propagating changes. The proposed model represents AI work as a directed acyclic graph (DAG) of artifact-producing computations with explicit dependencies and identity-based replay. In controlled experiments involving policy-memo updates, DAG replay demonstrated superior performance compared to loop-centric baselines. It achieved zero churn and prevented unrelated-branch contamination during updates, while ensuring perfect upstream preservation and downstream propagation during intermediate edits. The study highlights that final answer quality is distinct from maintained-state quality, noting that while loop baselines can produce polished outputs, they may mask state inconsistencies that compound over time. Execution lineage offers stronger guarantees for stable evolution of AI-generated work, addressing critical challenges in isolating irrelevant updates and managing complex, iterative AI tasks.
Wire timeline
From Agent Loops to Deterministic Graphs: Execution Lineage for Reproducible AI-Native Work
Researchers Josh and Seth Rosen introduce 'execution lineage,' a new execution model designed to enhance the reproducibility and maintainability of AI-native workflows. Current large language model systems often rely on implicit conversational states within agentic loops, leading to instability when preserving work products or propagating changes. The proposed model represents AI work as a directed acyclic graph (DAG) of artifact-producing computations with explicit dependencies and identity-based replay. In controlled experiments involving policy-memo updates, DAG replay demonstrated superior performance compared to loop-centric baselines. It achieved zero churn and prevented unrelated-branch contamination during updates, while ensuring perfect upstream preservation and downstream propagation during intermediate edits. The study highlights that final answer quality is distinct from maintained-state quality, noting that while loop baselines can produce polished outputs, they may mask state inconsistencies that compound over time. Execution lineage offers stronger guarantees for stable evolution of AI-generated work, addressing critical challenges in isolating irrelevant updates and managing complex, iterative AI tasks.
cs.AI updates on arXiv.org