Researchers from the École Polytechnique Fédérale de Lausanne (EPFL) in Switzerland have developed a new framework aimed at enhancing the efficiency of complex workflows involving large language models (LLMs). The team, which includes Sami Abuzakuk, Anne-Marie Kermarrec, and Rishi Sharma, introduced a system known as Workflow Optimisation (AWO). This innovative approach seeks to eliminate redundant steps in agentic workflows that typically rely on iterative reasoning and numerous interactions with tools, a process that can often lead to high operational costs and latency.
AWO works by analysing workflow data and consolidating repeated sequences of tool calls into what are termed “meta-tools.” This integration streamlines processes, thereby reducing dependency on LLM reasoning, which can sometimes produce errors. The researchers reported notable results, including reductions in LLM calls of up to 11.9% and increases in success rates by as much as 4.2 percentage points. The framework thus represents a significant advancement toward more efficient and reliable automation driven by LLMs.
The challenge of optimising agentic workflows is amplified by the high operational costs and potential for failure linked to repetitive reasoning. AWO addresses this by meticulously analysing existing workflow traces to identify common sequences of tool calls, subsequently transforming those sequences into reusable meta-tools. The process enables a more efficient execution path, effectively minimising the likelihood of errors while enhancing the reliability of these systems.
Significantly, the research highlights that many current tool sets are not ideally suited for agentic tasks, which creates opportunities for optimization through the deduplication of LLM efforts. The study indicated that many workflows exhibit a regular structure, particularly in their early stages, with a substantial number of tasks following similar trajectories. For instance, within the APPWORLD benchmark, at least 14.3% of tasks adhered to the same sequence after five steps. This consistent pattern inspired the creation of AWO, aiming to leverage this regularity to reduce inference costs and latency without sacrificing the inherent flexibility of agentic systems.
Technical Details
The static extraction of meta-tools from workflow trajectories prior to deployment is a key component of AWO. This method contrasts with runtime optimisation techniques and eliminates the associated overhead, ensuring deterministic execution of sub-tasks. By drawing on concepts similar to function inlining in compilers and kernel fusion in GPU programming, the researchers were able to adapt traditional optimisation techniques to workflows that involve non-deterministic control flows.
To develop meta-tools, the team conducted a detailed examination of workflow data to identify repetitive patterns amenable to abstraction. This process resulted in higher-level actions that alleviate the decision-making burden on the agent, leading to streamlined workflows. Performance gains were meticulously measured by comparing workflows executed with and without AWO-generated meta-tools, showcasing the framework’s potential to enhance the practicality and reliability of agentic systems in real-world applications.
The results indicate that integrating these meta-tools dramatically reduces both the number of LLM reasoning steps and total tool calls necessary for task completion. Researchers recorded a reduction in LLM calls across tested scenarios, reaching as high as 11.9%. Furthermore, AWO successfully improved task success rates by 4.2 percentage points. The team has made AWO available as an open-source framework to encourage community collaboration and further research in the optimisation of agentic workflows.
In essence, AWO addresses significant issues related to operational costs, latency, and hallucinations that stem from the iterative nature of reasoning and tool interactions in complex tasks. The framework not only streamlines these workflows but also maintains the flexibility essential for effective agentic systems. The researchers anticipate that future efforts could extend AWO’s capabilities to handle more intricate workflows and integrate it with additional optimisation techniques, which could further enhance the performance of agentic AI systems.
See also
Anthropic Launches Legal AI Tool, Sparks 10% Plunge in Software Stocks
AI Tools Revolutionize Work: Gemini, Midjourney, and Notion AI Transform Productivity and Creativity
Fibr AI Secures $7.5M Funding to Enhance Adaptive Web Personalization for Enterprises
