Workflow Optimization for LLM Agents

Analysis

Why This Paper Matters

LLM-based agents are increasingly used to solve complex tasks by composing workflows that interleave LLM calls, retrieval, tool use, and code execution. However, the field lacks a unified vocabulary and framework to describe and compare different optimization approaches. This survey addresses that gap by introducing the concept of agentic computation graphs (ACGs) and organizing the literature based on when workflow structure is determined—static vs. dynamic. This lens is crucial because it highlights a fundamental design choice: whether to fix a reusable scaffold before deployment or adapt the workflow per run.

The paper also clarifies the distinction between reusable workflow templates, run-specific realized graphs, and execution traces. This separation is important for understanding what is being optimized and what is actually deployed. By providing a structured taxonomy, the survey enables researchers to position new methods more clearly and compare existing work more effectively. This is particularly timely as the number of LLM agent frameworks and optimization techniques grows rapidly.

Technical Contributions

• Agentic Computation Graphs (ACGs): Proposes a unified abstraction for LLM-based workflows, treating them as graphs with components (agents) and dependencies.
• Taxonomy along three dimensions: (1) when structure is determined (static vs. dynamic), (2) what part of the workflow is optimized (e.g., components, dependencies, information flow), and (3) which evaluation signals guide optimization (task metrics, verifier signals, preferences, trace-derived feedback).
• Separation of templates, realized graphs, and traces: Distinguishes reusable design choices from run-specific structures and runtime behavior, enabling clearer analysis.
• Structure-aware evaluation perspective: Proposes complementing downstream task metrics with graph-level properties (e.g., complexity, robustness), execution cost, and structural variation across inputs.

Results

As a survey, this paper does not present new experimental results. Instead, its main contribution is a comprehensive framework for categorizing and comparing existing methods. It provides a clear vocabulary and a unified framework that can be used to position new methods and evaluate them more reproducibly. The paper does not include quantitative comparisons or benchmarks, but it sets the stage for future empirical studies by offering a structured way to describe and evaluate workflow optimization techniques.

Significance

The survey addresses a critical need in the rapidly evolving field of LLM agents: a common language and framework for discussing workflow optimization. By introducing ACGs and a multi-dimensional taxonomy, it helps researchers and practitioners understand the design space and trade-offs. The proposed structure-aware evaluation perspective encourages more holistic assessment beyond task accuracy, considering cost, robustness, and structural variation. This can lead to more reproducible and comparable research, accelerating progress in building efficient and reliable LLM-based systems. The framework is likely to become a reference point for future work in this area.