We’re Bullish on Agents

Chaining LLMs together to form teams of agents is a relatively new concept—it’s rapidly been adopted by big tech giants Google, Microsoft, Meta, and OpenAI. Adaptive agents aim to make this technology accessible to everyone who wants to explore it—read on to find out how we aim to do it!  

According to Andrew Ng, founder of DeepLearning.AI, agentic workflows are a novel way of interacting with and utilizing large language models (LLMs) that significantly enhance their capabilities and potential

Iterative and Interactive: Unlike traditional “one-shot” prompts where an LLM provides a single output in response to a user’s query, agentic workflows involve an ongoing back-and-forth dialogue between the user and the LLM. This facilitates a more nuanced and precise understanding of the user’s intent and enables the LLM to refine its responses accordingly.

Self-Reflection and Planning: Agentic workflows incorporate mechanisms that enable the LLM to critically assess its own outputs, identify potential limitations or errors, and generate plans for improvement. This iterative process allows the LLM to learn and enhance its performance over time.  

Tool Use: Agentic workflows allow LLMs to access and leverage a variety of external tools and resources, such as search engines, calculators, or code execution environments. This significantly expands the LLM’s capabilities, empowering it to tackle complex tasks that require external knowledge or computational capabilities.  

Multi-Agent Collaboration: In some cases, agentic workflows may involve the coordination and collaboration of multiple LLMs, each with specialized skills or knowledge domains. This enables the system to approach complex problems from multiple perspectives and generate more comprehensive and innovative solutions.

In summary, Andrew Ng sees agentic workflows as a groundbreaking development in the field of AI, offering a pathway to more powerful and versatile LLM-based applications. This approach not only enhances the capabilities of current LLMs but also lays the foundation for future advancements in artificial general intelligence (AGI).