MLOps.community A Candid Conversation Around MCP and A2A // Rahul Parundekar and Sam Partee // #316 SF Live
126 snips
May 21, 2025 Rahul Parundekar, Founder of AI Hero, Inc., and Sam Partee, CTO of Arcade AI, dive into the complexities of AI agents and tools. They tackle the significance of digital permissions and the challenges of agent-to-agent interactions. The duo unveils the intricacies of authentication processes, emphasizing OAuth's role in security. They also discuss the evolution of agent-based programming and AI tools, highlighting the need for improved evaluation methods. With humor, they address the frustrations of automated email responses while celebrating AI's potential to transform workflows.
AI Snips
Chapters
Transcript
Episode notes
Agents Extend LLM Functionality
- Agents connect language models with deterministic code to access external tools and data sources.
- This capability vastly expands the functionality of LLMs beyond memory and reasoning alone.
Authorization Limits Agent Actions
- Current systems lack proper authorization to let agents act on behalf of people securely.
- Elevating agent privileges without safeguards risks uncontrolled actions and potential damage.
Use Delegated Agent Privileges
- Implement delegated privileges for agents instead of sharing full user tokens.
- Use tool authorization separate from site authorization to avoid excessive access.