Optimizing Agentic Workflows using Meta-tools

TL;DR

The paper tackles the high latency and cost of agentic AI workflows caused by iterative reasoning and fine-grained tool usage. It introduces Agent Workflow Optimization (AWO), a framework that analyzes execution traces to identify recurring tool-call patterns and compiles them into deterministic meta-tools, implemented via state-graph representations with horizontal and vertical merging. Across two benchmarks, AWO reduces LLM calls by up to $11.9\%$ and improves task success by up to $4.2$ percentage points, while also reducing tokens and monetary cost; results vary by model and task complexity. The work demonstrates a practical path to making agentic systems more efficient and robust, and provides open-source tooling to enable reproducibility and further research in meta-tool-based optimization.

Abstract

Agentic AI enables LLM to dynamically reason, plan, and interact with tools to solve complex tasks. However, agentic workflows often require many iterative reasoning steps and tool invocations, leading to significant operational expense, end-to-end latency and failures due to hallucinations. This work introduces Agent Workflow Optimization (AWO), a framework that identifies and optimizes redundant tool execution patterns to improve the efficiency and robustness of agentic workflows. AWO analyzes existing workflow traces to discover recurring sequences of tool calls and transforms them into meta-tools, which are deterministic, composite tools that bundle multiple agent actions into a single invocation. Meta-tools bypass unnecessary intermediate LLM reasoning steps and reduce operational cost while also shortening execution paths, leading to fewer failures. Experiments on two agentic AI benchmarks show that AWO reduces the number of LLM calls up to 11.9% while also increasing the task success rate by up to 4.2 percent points.

Figures (10)

• Figure 1: Percentage of tasks where agents have equivalent trajectories at a given step in AppWorld (lower bound).
• Figure 2: The ReAct Loop. The agent interleaves reasoning with external execution ("Action"), creating a step-by-step workflow.
• Figure 3: (top) Two distinct user prompts (in blue) and the associated workflow, including LLM reasoning and tool calling. User prompts, while different from a surface-level, might share sequences of tool calls. Our framework leverages this property to create meta-tools, reducing the number of costly LLM calls.
• Figure 4: Overview of the AWO framework. Existing agent workflows are first transformed into a state graph (step 1) and then compressed via horizontal merging into a merged state graph to expose redundant workflows that end up in equivalent states (step 2). The merged state graph is used to identify meta-tools (step 3) which are added to the agent’s tool set (step 4).
• Figure 5: Meta-Tool extraction from a merged state graph