I recently attended the Agentic AI Summit 2025 at UC Berkeley, and I’m still processing the wealth of information shared by experts in the field. As someone who’s been following the development of agentic AI, I was excited to see the progress made and the challenges that still lie ahead.
The summit drew a large crowd, with over 1,500 attendees and a significant online presence. The talks were technical and to the point, which I appreciated. A common theme throughout the summit was that the main bottleneck in agentic AI development isn’t training large models, but rather steering and managing them in real-world systems.
Several projects showcased promising approaches to agentic AI. ReAct-style feedback loops, for instance, allow large language models to reason, ask for outside help, and repeat the process. The Model Context Protocol (MCP) enables different agents and tools to communicate with each other in a more modular way. Memory, however, remains a significant challenge, with no one seemingly satisfied with the current solutions.
The summit also highlighted the importance of frameworks and standards in agentic AI development. CrewAI, LangGraph, LlamaIndex, and Goose were some of the frameworks mentioned, while Agntcy.org, FRAMES, and Mozilla’s open-source agent tools were introduced as new standards.
One of the most significant takeaways from the summit was that the reality of agentic AI doesn’t quite live up to the hype. While agents are making progress in tasks like form filling and coding, they still struggle with context and autonomy. The cost of running these systems is also a significant concern, with the fees adding up quickly.
Despite the challenges, the summit left me feeling optimistic about the future of agentic AI. The teams making the most progress are those focusing on context, logistics, and performance measurement. As the field continues to evolve, I’m excited to see how these challenges are addressed and what new breakthroughs are made.
If you’re interested in learning more, I recommend checking out the summit’s website, which has recordings of the sessions available. I’d also love to hear from others who are experimenting with agentic AI – what are some of the challenges you’re facing, and how are you overcoming them?