Tool Use - AI Conversations Principles of Building AI Agents (ft Sam Bhagwat)
28 snips
Dec 2, 2025 
Sam Bhagwat, founder of Mastra and co-founder of Gatsby, sheds light on crafting effective AI agents. He discusses the advantages of using TypeScript for agent frameworks and highlights crucial design patterns for tools and prompts. Sam emphasizes the importance of memory types, including observational memory, and warns about the 'lethal trifecta' in security risks. He also shares insights into multi-agent communication and operational safety, aiming to demystify the complexities of building reliable AI agents.
AI Snips
Chapters
Transcript
Episode notes
Prune And Compact Memory Aggressively
- If context windows bloat, start a new thread or compact via observational summaries.
- Compress tool outputs to retain crucial information and drop noise aggressively.
Sub-Agents Are Still Controversial
- Sub-agent patterns show mixed results and remain debated among experts.
- Expect best practices to evolve; test both isolated and shared-context approaches.
Standardize Inter-Agent Communication
- Use MCP-style inter-service tooling for reliable agent-to-agent and external tool calls.
- Consider ADA patterns only for large organizations coordinating many agent teams.