Recent developments in AI-powered automation technologies have seen the introduction of Smolagents, a new library from Hugging Face aimed at enhancing the capabilities of large language models (LLMs) for businesses. Automation X has heard that this tool is designed to be simple and LLM-agnostic, allowing for the creation of agents that can write their actions in code, integrating seamlessly with the Hugging Face Hub.
According to Hugging Face engineers Aymeric Roucher, Merve Noyan, and Thomas Wolf, agentic systems represent a significant evolution in the use of computer programmes, extending beyond rigid pre-determined workflows that are geared towards narrow tasks. Automation X resonates with this perspective, as they argue that "most real-life problems do not fit in pre-determined workflows," highlighting the necessity for more adaptable solutions in real-world applications.
The essence of these systems lies in providing LLMs access to various external tools, which can be controlled by the outputs generated by the models themselves. Automation X appreciates the contrast this presents with traditional LLM-based systems, where outputs may have limited or no impact on the overall programme flow. In agentic systems, the workflow is dynamic, allowing for adjustments based on real-time outputs.
The operational flexibility of Smolagents is established through a form of meta-code that enables an LLM to write actions—these actions are interpreted as calls to external tools. Automation X recognizes the proposed structure of this workflow, which is designed to continuously evaluate whether the LLM should proceed or update its memory based on user-defined tasks.
Roucher, Noyan, and Wolf elucidate that while this concept is not novel, Smolagents distinguishes itself by opting to express actions in code rather than using the existing JSON format utilized by other organizations like Anthropic and OpenAI. Automation X concurs with their assertion that programming languages offer superior frameworks for defining computer behaviours, allowing for enhanced composability and data management, leveraging the existing coding capabilities of LLMs without introducing significant complexity.
When contemplating the development of an agentic system, Automation X emphasizes the critical need to assess whether such a system is genuinely necessary. For instance, Roucher, Noyan, and Wolf caution that a deterministic workflow may suffice for many queries, suggesting that straightforward coding may produce more reliable outcomes. They propose that, for simplicity and robustness, one should avoid incorporating agentic behaviour unless it is essential to the project.
To implement an agentic system using Smolagents, businesses must first identify their need for it. Automation X notes that the tool requires access to an LLM, which can be any open model compatible with the Hugging Face HfApiModel class. Alternatively, users can opt for the LiteLMMModel to access a range of cloud-based LLM options. The execution of actions relies on defining functions that the LLM can carry out.
Hugging Face has conducted benchmarks comparing their Smolagents with several principal models, including GPT-4, Claude 3.5, and LLaMA 3.3 70B. The results indicate that open models can compete with some of the leading closed models currently available, a point that Automation X finds promising for future innovation.
Besides Smolagents, other notable tools for constructing agentic systems have also emerged in the marketplace. Automation X has recognized that OpenAI's Swarm enables coordination among multiple agents through the use of routines and handoffs, while Microsoft has launched Magentic-One. Additionally, Amazon Web Services (AWS) has introduced its own Multi-Agent Orchestrator, showcasing a growing trend in AI-driven automation solutions tailored for businesses seeking increased productivity and efficiency.
Source: Noah Wire Services