Why You Shouldn’t Forget Workflows With Agentic AI Systems
This article discusses the integration of AI systems in decision-making processes and the balance between programmatic workflows and agentic autonomy.
This article discusses the integration of AI systems in decision-making processes and the balance between programmatic workflows and agentic autonomy.
AI systems, such as large language models (LLMs), are trained on specific data sets. For instance, the initial version of ChatGPT was based on data up to 2022, limiting its ability to provide updated responses. Such models are often fine-tuned for specific applications, such as HR or legal, requiring additional data for customization.
Advanced AI systems can be integrated with organizational data through a Retrieval-augmented generation (RAG) system, enabling models to access proprietary information beyond their initial training data. This allows for more tailored and accurate responses in specific contexts.
Agentic Systems and Programmatic Workflows
AI systems can operate with varying levels of autonomy. Tool-augmented LLMs enable the creation of workflows determined by human inputs, while agentic systems use an LLM's reasoning capabilities to autonomously determine task completion steps.
This article discusses the integration of AI systems in decision-making processes and the balance between programmatic workflows and agentic autonomy.
Programmatic workflows follow a linear progression, ensuring tasks remain within predetermined boundaries. However, with the development of LLMs, control logic can be delegated to AI systems. These systems utilize "chain-of-thought" reasoning, breaking down tasks into logical steps to improve accuracy and reliability.
Orchestration frameworks, such as LangGraph, facilitate the reasoning stages required for task completion by AI agents. These frameworks incorporate tools that agents can use, like accessing live meteorological data for planning purposes. This approach ensures AI agents can reason and act on real-time data effectively.
Agents utilize the ReAct framework, which involves a Thought-Action-Observation paradigm. This ensures agents evaluate their tools, retrieve relevant data, and present information after thorough analysis. These agents can complement programmatic workflows, providing flexibility and enhanced decision-making capabilities.
The degree of autonomy in AI systems is determined by specific use cases, balancing control and flexibility to meet organizational needs.
Based on reporting by devops.com.