Workflow Optimization Achieves 11.9% Gains in Agentic Task Efficiency

Researchers are tackling the challenge of optimising complex tasks assigned to large language models (LLMs), which frequently rely on iterative reasoning and numerous tool interactions. Sami Abuzakuk, Anne-Marie Kermarrec, and Rishi Sharma, all from EPFL, Switzerland, alongside colleagues Veski et al., present a novel framework called Workflow Optimisation (AWO) designed to identify and eliminate redundant steps within these agentic workflows. AWO achieves this by analysing workflow data and consolidating recurring tool call sequences into ‘meta-tools’, effectively streamlining processes and reducing reliance on potentially error-prone LLM reasoning. This innovation demonstrably lowers operational costs and latency, while also improving task success rates , experiments reveal reductions in LLM calls of up to 11.9% and success rate increases of up to 4.2 percentage points, representing a significant step towards more efficient and reliable LLM-driven automation.

AWO achieves this by analysing workflow data and consolidating recurring tool call sequences into ‘meta-tools’, effectively streamlining processes and reducing reliance on potentially error-prone LLM reasoning.

Optimising LLM agents via meta-tools requires careful orchestration

These agents, capable of dynamic reasoning and interaction with tools, often suffer from high operational costs, latency, and failures due to repetitive reasoning steps and potential Hallucinations. AWO meticulously analyses existing workflow traces to pinpoint recurring sequences of tool calls, subsequently transforming them into ‘meta-tools’. The team achieved substantial improvements by creating these meta-tools, which shorten execution paths and minimise the likelihood of failures. This demonstrates that streamlining workflows not only lowers computational costs but also enhances the reliability of agentic systems.

This work establishes a method for analysing agentic workflows and detecting common, repetitive patterns in tool usage. By compiling these patterns into meta-tools, AWO allows agents to bypass redundant reasoning steps, leading to faster and more cost-effective task completion. The research highlights that current tool sets are not always optimally designed for agentic use, creating opportunities for significant gains through deduplication of LLM effort. The framework provides a means to analyse tool call trajectories and detect recurring sequences, encapsulating common behaviours into single, callable operations.

Researchers observed that agentic workflows frequently exhibit regular structure, particularly in their initial stages, with a significant proportion of tasks following identical trajectories at certain steps. For example, after five steps in the APPWORLD benchmark, at least 14.3 percent of tasks followed the same sequence of actions. This observation motivated the development of AWO, which aims to capitalise on this regularity to reduce inference costs and latency without compromising the flexibility inherent in agentic systems. The study unveils a practical approach to optimising agentic AI, paving the way for wider adoption, improved scalability, and enhanced reliability in increasingly complex environments.

Static Meta-tool Extraction for LLM Workflows enables improved

This innovative approach bypasses unnecessary intermediate LLM reasoning steps, directly reducing operational costs and shortening overall execution times. The core of the work involved static extraction of these meta-tools from workflow trajectories prior to deployment. This contrasts with runtime optimisation techniques, eliminating associated overhead and guaranteeing deterministic execution of sub-tasks. Scientists harnessed principles analogous to function inlining in compilers and kernel fusion in GPU programming to inform the design of AWO. The team analysed agent trajectories, identifying common tool sequences and fusing them into meta-tools, adapting traditional compiler optimisation to accommodate non-deterministic control flows.

This process involved detailed examination of workflow data to pinpoint repetitive patterns suitable for abstraction. The resulting meta-tools function as higher-level actions, effectively reducing the agent’s decision load and streamlining the workflow. The research team meticulously measured performance gains by comparing workflows executed with and without AWO-generated meta-tools. This precise measurement approach enabled a clear demonstration of AWO’s benefits, showcasing its potential to improve the practicality and reliability of agentic systems in production environments. The technique reveals a complementary relationship with other optimisation strategies, such as parallel execution and improved search algorithms, suggesting potential for further enhancements.

Meta-tools reduce LLM calls and latency, improving efficiency

The research addresses the significant operational expense, latency, and potential for hallucinations that arise from iterative reasoning and tool invocations in complex tasks. This bypasses unnecessary LLM reasoning steps, reducing operational costs and shortening execution paths. These results confirm that the framework not only lowers inference costs and latency but also maintains the generality crucial for effective agentic systems. The study involved analysing large collections of execution traces generated by state-of-the-art LLM agents, identifying patterns suitable for compilation into meta-tools across API-driven application workflows and interactive web-based tasks.

Data shows that integrating these meta-tools substantially reduces both LLM reasoning steps and the total number of tool calls required for task completion. Specifically, the reduction in LLM calls reached as high as 11.9 percent across the tested scenarios. The work introduces a novel framework for detecting redundant patterns in agentic workflows through merging agentic executions, allowing for the creation of meta-tools that minimise unnecessary reasoning and token usage. Researchers thoroughly evaluated AWO on a diverse set of agentic workflows, demonstrating its ability to reduce token usage by up to 11.9 percent and improve task success rates by 4.2 percentage points. The team has released AWO as an open-source framework to facilitate reproducibility, encourage community collaboration, and stimulate further research in agent optimisation. The study highlights how existing workflows can be made more efficient by coalescing tool calls, as demonstrated through examples involving tasks such as sending documents and summarising performance reviews.

Meta-tools streamline agentic AI workflow efficiency by automating

These workflows, which involve dynamic reasoning, planning, and tool use, often suffer from high operational costs, latency, and potential failures due to inconsistencies. AWO addresses these issues by identifying and optimizing redundant patterns of tool execution within existing workflow data. By bypassing unnecessary intermediate reasoning steps performed by large language models, AWO reduces computational expense and execution time, ultimately leading to improved task success rates. This work draws parallels with kernel fusion techniques used in GPU computing, where multiple operations are combined to reduce memory traffic and improve efficiency.

AWO adapts these principles to agentic workflows, effectively compiling common agent behaviours into reusable components. The authors acknowledge that their framework operates within the constraints of stochastic, LLM-generated execution traces, which introduces inherent limitations. Future research could explore extending AWO to handle more complex workflow scenarios and integrating it with other optimization techniques to further enhance the performance of agentic systems.