Session brief
Keynote: Why people think \"agent\" is a buzzword but it isn't
Overview
This talk challenges the perception of AI agents as mere buzzwords, arguing instead for their significant potential. The speaker defines agents as entities that can perceive and act upon their environment, drawing parallels from historical definitions to modern applications like coding assistants. The core thesis is that while agents are not new, building and deploying them effectively presents substantial challenges that, once overcome, will unlock transformative use cases.
Who should watch
- AI engineers and developers building agentic systems.
- Product Managers exploring the capabilities and limitations of AI agents.
- Builders seeking to integrate AI into existing workflows and applications.
- Anyone interested in the practical challenges and future potential of AI agents beyond the hype.
Key takeaways
- Agents are defined as anything that can perceive and act on its environment, a concept with roots in AI literature from the 1980s and 90s.
- Actions allow models to overcome limitations like outdated knowledge or poor mathematical abilities by integrating tools like web browsers or calculators.
- The primary challenge in agent development is the "curse of complexity," where task failure rates increase exponentially with the number of steps required.
- Many current agents struggle with tasks requiring more than five sequential steps, highlighting a need for improved planning and reasoning capabilities.
- "Tool use" (natural language to API translation) is another significant hurdle, stemming from ambiguous natural language inputs and poorly documented or designed APIs.
- Effective tool use requires detailed documentation, including function descriptions, parameter types, error codes, and expected return values, with clear guidance on error handling.
- Context management is crucial, as agents require significant information, including tool documentation, execution history, and reasoning steps, which can quickly exceed model context windows.
- Models good at planning are not always good with long contexts, and vice versa, suggesting a need for research into models that excel at both.
Notable quotes
*A definition that's like really resonate with me is from the book by Stuart Russell and Peter Norvig; they defined an agent as anything that can perceive the environment and that act on the environment.*
*Doing agents is like really really hard.*
*The success rate the pass rate like decrease rapidly as the number of steps increase.*
Unofficial community note. Prefer the recording for nuance.