From Agent Loops to Deterministic Graphs: How Execution Lineage Makes AI-Generated Work Maintainable

This article introduces execution lineage, a new execution model designed to address the fundamental flaws of traditional Agent loops in state management, change propagation, and reproducibility. The model represents AI-native work as a directed acyclic graph (DAG) with explicit dependencies, enhancing the maintainability and reproducibility of AI-generated work through features like explicit dependencies, stable intermediate boundaries, and identity-based replay.

The Agent loop workflow of large language model systems (reasoning, tool calling, memory access, iterative optimization) seems efficient on the surface, but it has implicit dialogue state issues: when modifications to intermediate steps, change isolation, or decision tracing are needed, the system often has to regenerate large amounts of content, even introducing side effects; the same prompt may produce different results at different times, leading to a lack of reproducibility.

Execution lineage represents AI-native work as a directed acyclic graph (DAG), where nodes stand for computation steps and edges represent explicit dependencies. Its key features include:

1. Explicit Dependencies: Each step clearly specifies its input sources, laying the foundation for change tracking;
2. Stable Intermediate Boundaries: Intermediate products have unique identifiers and are stored stably, decomposed into independently manageable modules;
3. Identity-Based Replay: Precisely locates input states using identifiers to ensure the same input produces the same output.

Comparative experiments validate the value of execution lineage:

• Scenario 1 (Irrelevant Branch Update): DAG replay fully preserves the original state and only modifies the target node and its downstream nodes; the loop baseline regenerates the entire content, which easily introduces irrelevant information leading to drift.
• Scenario 2 (Intermediate Product Editing): DAG replay outperforms the loop method in four metrics—upstream preservation, downstream propagation, protection of unaffected products, and cross-product consistency—avoiding content omissions or contradictions.

The study reveals that final answer quality and maintainable state quality are distinct dimensions. A strong loop baseline can produce high-quality final outputs, but it may mask some state inconsistency issues, which will accumulate and amplify in future revisions. Just as a writer needs to explain the source of paragraphs and modify arguments without breaking the structure, AI systems need to organize their work processes structurally to become maintainable knowledge assets.

Execution lineage provides practical guidance for AI-native applications:

• Paradigm Shift in Version Control: Precisely answers the scope of impact of modifications for differences in computation processes;
• Collaboration Support: Clear boundaries reduce collaboration conflicts;
• Audit and Compliance Foundation: Traces the complete lineage of decisions;
• Incremental Computation Optimization: Reduces unnecessary computations and improves response speed.

Limitations of execution lineage: Implementation requires explicit modeling, increasing the development burden in the rapid prototyping phase; it may constrain thinking for highly creative tasks. Future research directions: Automatically extracting lineage structures from Agent loops, reducing modeling costs, and integrating with existing MLOps/LLMOps toolchains.

The shift from Agent loops to deterministic graphs is not only an evolution of technical architecture but also a change in expectations for AI systems—from 'giving good answers' to 'continuously producing maintainable work'. Execution lineage provides a theoretical foundation and practical path; in today's era of widespread AI-native applications, the pursuit of maintainability will determine whether a system can become a production-grade tool.