| name | llm-agent-workflow-optimization-survey |
| title | From Static Templates to Dynamic Runtime Graphs: A Survey of LLM Agent Workflow Optimization |
| version | 0.0.3 |
| engine | skillxiv-v0.0.3-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2603.22386 |
| keywords | ["LLM Agents","Workflow Optimization","Agentic Computation Graphs","Dynamic Planning","Structure-Aware Evaluation"] |
| description | Navigate LLM agent workflow design by understanding the taxonomy of static vs dynamic methods and the dimensions that organize them. Agentic Computation Graphs (ACGs) framework distinguishes when structure is determined (before vs during execution), which components optimize, and what signals guide optimization. Provides structure-aware evaluation criteria combining downstream metrics with graph properties and robustness. Use when designing or optimizing agent workflows, choosing between fixed templates and dynamic adaptation, or evaluating workflow efficiency. |
| category | Survey & Synthesis |
Field Overview
LLM agent workflows—the orchestration of language model reasoning, tool calling, and feedback loops—have become a central research area. However, the field lacks a common taxonomy for describing and comparing workflows. Papers discuss "planning," "prompting," "tool selection," and "re-planning" but with inconsistent terminology. This survey unifies the landscape around Agentic Computation Graphs (ACGs): abstract structures representing how an agent decides to call tools, in what order, and with what parameters.
Taxonomy: The ACG Framework
The survey organizes workflows along three orthogonal dimensions:
Dimension 1: When Structure Is Determined
- Static workflows: Fix a reusable scaffold before deployment. Examples: ReAct-style fixed chains, predefined tool sequences. Advantage: reproducible, easy to optimize. Disadvantage: rigid, doesn't adapt to task variation.
- Dynamic workflows: Select, generate, or revise the workflow for a particular run. Happens before execution (planning) or during execution (reactive). Advantage: adapts to task; Disadvantage: higher latency, harder to optimize, less reproducible.
Dimension 2: Which Component Optimizes
Each workflow stage (reasoning, tool selection, tool sequencing, execution feedback) is an optimization point:
- Reasoning optimization: Better prompts, better in-context examples, chain-of-thought structures
- Tool selection: Which tools are available for this task? Should this task call tools at all?
- Tool sequencing: In what order should tools be called? Parallel or serial?
- Execution feedback: How does the agent learn from tool results?
Dimension 3: What Guides Optimization
Four signal sources tell the agent whether the workflow is working:
- Task metrics: Final answer accuracy, downstream performance
- Verifier signals: Does an external model think the answer is correct?
- Preference signals: Human feedback on which workflows are better
- Trace feedback: Properties of the workflow itself (execution cost, token usage, intermediate reasoning quality)
Key Distinctions the Survey Makes
The survey clarifies three often-confused concepts:
Workflow Template vs Realized Graph vs Execution Trace
- A template is the reusable blueprint (e.g., "if task is math, use calculator; if task is reading, use search"). Shared across many problem instances.
- A is the specific workflow instantiated for one problem (e.g., "for this math problem, call calculator once then reason"). Generated from the template.
