Agent AI Summit Note @ Berkeley

My takeaway:
In a world transformed by large language models (LLMs) to come, a new reality is emerging: a world full of AI agents that can communicate and collaborate on complex tasks. This could be how AI seamlessly integrates into our daily lives.
Key technologies driving this shift will likely be Agent-AI SDKs, Protocol (e.g., Model Context Protocol (MCP), A2A) for communication among agent, tools and data. Services for AI (e.g., MCP server) could receive a significant boost from these advancements.

Link to all videos at the end.        

1, Hardware for AI Agents
Bill Dally / NVIDIA
– El Agente: demo quantum chemistry
– LLM —  Agent: state, memory, policy – tools
– 3D parallelism (tensor, data and pipeline parallel)
– Decode phase reads entire model and KV cache for each token
– challenges: 
  disaggregated inference;  chain/tree of thought (high token rates), evolving model


2, NLIP
 Ranjan Sinha (IBM)
– NLIP, or the Natural Language Interaction Protocol, is a proposed protocol designed to standardize communication between intelligent agents, particularly those utilizing natural language

        

3, Architecting the Future of Agentic AI at Scale
Ramine Roane
– inference : core workload for agentic AI


4, vLLM:
– how to handle multiple current request to LLM, each needs a pass through the model
– static batching and dynamic batching
– Paged Attention: storing continuous KV in non-contiguous memory space, like OS
– 

5, Building AI Infrastructure as If Agents Were Human
Chuan Li
– Agent communication,  RPC < 50ms, 
– 

6, MCP as a Foundational Protocol for Agents"
Joson Kim
–  Models are only as good as the context provided to them
–  protocol between AI apps & agents interact with tools and data sources
        

7, Ember: the inference-time scaling architectures framework
– Torch::NNs::Ember:NoN
–  compound AI systems,



8, A2A: The Future of AI Agent Collaboration
– Modality 

9, LMArena and MAST

Productizing a feature takes 10X-50X efforts
LLM: really have clear specification – hard to know an error
ChatBot Arena/LMArena: human evaluation for LLMs
   Most successful AI apps have human in the loop
   Human preference: substance and styles ( stype: how human interact human)
   Quantifying human preference
MAST: multi-agent system failure taxonomy
  50% of fail
  
10, Reflective Optimization of Agents with GEPA and DSPy
– how to teach AI new tasks?
– weight update is the existing way (FLOPS get cheaper, sampling)
– RL with verified rewards
– GEPA (vs GRPO): use traces of rollouts and update the prompt

        

11, The Year of Agents (tm)
- why agents is not everywhere: connectors, generalization, hard to create agents, cost
    Connectors: MCP, internal knowledge 
    Generalization: agent tools, not good at custom tools
    
12, Visionary Stacks for Agentic Systems: Insights and Innovations Ahead
– NorthStar: AI Agent that grows with you
Natural Interface – Computers with Senses
Challenges: consensus, shared context, interop of mange AI models

13, Context Engineering and MCP for Document Workflows


14, Towards Building Safe and Secure Agentic AI
Agentic AI makes AI more valuable as it has data/memory, tools, potential communication involved.
AgentBeats, PromptPosion, AgentVigil
Use AI to build security tools. 

15, Google Gemini Era: Bringing AI to Universal Assistant and the Real World
Project Astra: agents and reasoning capability that solve complex problem


        

All talks and videos are available:
https://rdi.berkeley.edu/events/agentic-ai-summit