Welcome to Part 2 of the 3-Part Series on Agentic Workflows and AI Agents. In Part 1, we explored the rise of AI agents and the technological breakthroughs that are aiding its rise in capabilities. To recap, AI agents are autonomous/semi-autonomous systems designed to accomplish specific and complex tasks. In this edition of the newsletter, we will be exploring the rise of design patterns that influence the development of building effective AI agents.
Design Patterns — What are they?
Design patterns in AI agents are like tried-and-true recipes for solving common problems in artificial intelligence. Just as a chef might use a familiar recipe to create a delicious dish, AI developers use these design patterns to build smarter, more efficient software agents that can tackle complex tasks and make decisions in various situations.
Drawing from the insights of AI pioneer Andrew Ng, we can identify five key design patterns that give rise to the creation of AI agents:
1. Planning: AI agents use LLMs to devise and implement multi-step strategies to meet specific goals. This ability to break down complex tasks into manageable steps is crucial for tackling complex problems.
2. Tool Use: Agents are equipped with a variety of tools, such as web scrapping capabilities and code execution, enabling them to gather information, make decisions, and process data more effectively.
3. Data or Context Understanding: Many agents employ Retrieval-Augmented Generation (RAG) capabilities or digest specific datasets to enhance their task completion abilities. This allows them to use relevant information and context when devising solutions.
4. Reflection: Through the use of LLMs, agents can evaluate their own performance, identifying areas for improvement and refining their approaches over time.
5. Multi-Agent Collaboration: In more advanced systems, multiple AI agents can work together, distributing tasks and exchanging ideas to create superior solutions.
The patterns described above serve as foundational frameworks for developing AI agents that are both robust and adaptable. By leveraging these design patterns, AI developers can create systems that not only perform specific tasks efficiently but also constantly learn and evolve over time.
The Symbiotic Relationship between Agents and LLMs
While LLMs excel at generating human-like text and understanding context, they struggle with specific tasks like logic, calculation, and search. AI agents bridge this gap by tapping on LLMs for their strengths while using specialized tools to compensate for their weaknesses.
For instance, when an agent encounters a problem requiring a specific skill outside the LLM’s capabilities, it can rely on an appropriate tool from its toolbox. This approach has led to excellent scores on several benchmarks, including impressive results on challenges like HumanEval.
“AI agents will become the primary way we interact with computers in the future. They will be able to understand our needs and preferences, and proactively help us with tasks and decision making”
– Satya Nadella, CEO of Microsoft
The Future of AI Agents
As we look to the future, the potential applications of AI agents are endless. From improving the efficiency of customer service interactions to streamlining scientific research, AI agents have the potential to transform numerous industries and fields of study.
One particularly promising area is the development of AI-originated applications. These applications are built from the ground up with AI agents at their core, enabling new levels of personalization and problem-solving capabilities.
Defined is working closely with our portfolio companies and supporting the product strategy and technology roadmaps to further agentic workflows. In particular, Arcanna is codifying agentic workflows into their AI autopilot to advance cybersecurity while Quandri is optimizing operations in insurance with AI Agents.
Stay tuned for Part 3, where we will examine blueprints for AI-native approaches for building applications that reimagine workflows from first principles.
If you are a founder or know of people actively exploring automation with agentic workflows and further the capability ladder of AI Agents, please reach out to connect.