Claude Agent Skills - Dynamic Context Retrieval + Roadmap ahead for Agentic Frameworks

Video Description

LLM Agents have proven to be versatile and can do a wide range of tasks. Most of these tasks require the use of specialised tools equipped to the LLMs. The LLM's main role is then the selection of the right tool based on the task. However, as the number of tools increases, and the tool usage grows in complexity, too much information is fed to the LLM Agent. Can we reduce the context given to the LLM Agent, and still perform the task? This is where Anthropic's Agent Skills comes into play: It provides only the name and description of the skill to the agent Agent can decide to load into context of a specific skill if needed Agent can use code to execute the skill This modularised dynamic loading of context is interesting, and can definitely help to mitigate issues of too much context. But wait... what if there is another method, where we just use agents minimally and instead use more rule-based pipelines for the heavylifting? Is the dogma that an agent should do everything from start to end still a viable one in today's context? ~~~ Reference documentation on Claude Skills: https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills Slides: https://github.com/tanchongmin/john-youtube/blob/main/Discussion_Sessions/Agent%20Skills/Agent%20Skills.pdf Sample Code I did for Agent Skills: https://github.com/tanchongmin/john-youtube/blob/main/Discussion_Sessions/Agent%20Skills/Agent%20Skills.ipynb ~~~ My open-sourced LLM structured output parser, strictjson: https://github.com/tanchongmin/strictjson My open-sourced agentic framework, agentjo: https://github.com/tanchongmin/agentjo ~~~ 0:00 Introduction 1:24 Traditional Agent + Tools 11:58 Code Agent 17:23 Dynamic Context 24:16 I know Kung Fu 25:00 How Claude Skills work 55:55 Code Demo using Output Parser 1:17:28 Multi-Agent Systems 1:24:11 Pipeline-based System 1:35:48 LLMs are not all you need 1:39:38 Discussion 2:03:17 Conclusion ~~~ AI and ML enthusiast. Likes to think about the essences behind breakthroughs of AI and explain it in a simple and relatable way. Also, I am an avid game creator. Discord: https://discord.gg/bzp87AHJy5 LinkedIn: https://www.linkedin.com/in/chong-min-tan-94652288/ Online AI blog: https://delvingintotech.wordpress.com/ Twitter: https://twitter.com/johntanchongmin Try out my games here: https://simmer.io/@chongmin