Agent Patterns Explained

Augmented LLMs (Large Language Models): These models are pivotal in enhancing basic LLM capabilities by integrating additional tools and memory. According to Anthropic, this setup represents their simplest workflow within the realm of agent patterns. It closely resembles what they consider a classic agent. Typically known as a classic RAG (Retrieval Augmented Generation) workflow, this process involves the LLM accessing a database—usually a vector database—to gather necessary data. The LLM may then use various tools to process this data, and it taps into its memory to consider past interactions before producing a response. This approach significantly improves the quality of interactions, providing responses that are not only precise but also contextually aware, making them more relevant and useful.

Prompt Chaining: Prompt chaining is an effective method for managing complex tasks by breaking them into sequential steps. This approach is crucial because large tasks can overwhelm AI models, leading to poor outcomes. For example, instead of asking an AI to build an entire e-commerce site, the task can be split into smaller parts, such as developing the backend or setting up microservices on AWS. This method allows for detailed, focused processing of each component, improving the AI’s performance and results. Anthropic recommends prompt chaining as a best practice for complex AI-driven projects

Routing: Routing is a compelling workflow pattern that plays a critical role in directing tasks efficiently within AI systems. In this setup, an LLM router acts as the decision-maker that determines which specific LLM should handle an incoming request.

This process involves several specialized LLMs, each configured to perform distinct tasks. They are equipped with their own system prompts and may have access to specialized tools or even different LLM configurations. For instance, the first LLM might be finely tuned for one type of task, while others might be optimized for different areas.

Parallelization: Parallelization is a key workflow pattern where a complex task is divided into smaller, more manageable segments and distributed across multiple LLMs. This division allows each segment of the task to be processed simultaneously, significantly enhancing the efficiency of operations. Unlike routing, where tasks are sent to different LLMs based on their specialization, parallelization focuses on executing multiple tasks concurrently to speed up the overall process. An aggregator function then collects and combines results from all segments, ensuring a cohesive output. This method is particularly valuable in scenarios where speed is crucial and tasks can be naturally segmented to operate in parallel, thus shortening response times and improving performance.

Orchestrator-Workers:The orchestrator-workers model enhances task management within AI systems. This pattern involves an orchestrator or supervisor that breaks down complex tasks into manageable units and assigns them to specialized worker agents. For example, in financial research, the orchestrator delegates specific economic data tasks—like GDP analysis or gas price monitoring—to agents skilled in those areas.

Each agent activates only when needed, improving efficiency by focusing on their expertise. After completing their tasks, the results are consolidated by a synthesizer node, which integrates the outputs into a cohesive final report. This system is particularly effective in scenarios requiring diverse and specialized knowledge, ensuring precise and efficient task execution.

Evaluator-Optimizer: The Evaluator Optimizer pattern involves a feedback loop where one AI model generates an output and another evaluates and refines it. For instance, in software development, a coder agent writes code that a reviewer agent evaluates for quality. The reviewer provides feedback, which the coder uses to enhance the code. This cycle repeats until the code meets quality standards, ensuring continuous improvement and precision in the outputs.

Autonomous Agents: Autonomous Agents independently handle complex, open-ended problems within AI systems, similar to the functionality seen in Microsoft’s Autogen framework. These agents autonomously perform actions based on real-time data and feedback, adapting their responses as needed without human oversight. For example, an autonomous agent tasked with creating a sign-up form for an online banking system might generate code, test it, and revise it based on the outcomes of those tests, continuing this cycle until the code meets the specified standards. This self-guided process allows for high adaptability but can also lead to unpredictability in performance and costs, requiring careful monitoring and management to ensure efficiency.