Execution Lineage: DAG-Based Reproducibility for AI-Native Workflows
A recent paper on arXiv (2605.06365) presents a novel execution model known as execution lineage, designed for large language model systems. This model conceptualizes AI-generated tasks as a directed acyclic graph (DAG) comprising computations that produce artifacts. Its purpose is to ensure that evolving AI outputs remain manageable amidst changes by establishing clear dependencies, stable intermediate boundaries, and identity-based replay. The researchers evaluated DAG replay against traditional loop-centric update methods in two controlled policy-memo update scenarios, discovering that DAG replay accurately maintained the final memo during an unrelated-branch update. This method tackles the challenge of implicit conversational states within agentic workflows that blend reasoning, tool usage, memory, and iterative refinement.
Key facts
- arXiv paper 2605.06365 proposes execution lineage for AI-native work
- Execution lineage represents work as a directed acyclic graph (DAG) of artifact-producing computations
- The model includes explicit dependencies, stable intermediate boundaries, and identity-based replay
- Goal is to make evolving AI-generated work maintainable under change
- Compared DAG replay against loop-centric update baselines on two policy-memo update tasks
- In an unrelated-branch update, DAG replay preserved the final memo exactly
- Addresses implicit conversational state in agentic workflows
- Workflows interleave reasoning, tool use, memory, and iterative refinement
Entities
Institutions
- arXiv