AI :: Agent :: Training :: Principles of Building AI Agents :: Principles of Building AI Agents | Sam Bhagwat, CEO Mastra
⪼ Made with 💜 by Polyglot.
This video is an interview-style talk with Sam, the author of "Principles of Building AI Agents" and a founder of MRA (a YC startup). The intent is to educate and inspire viewers about building AI agents, sharing foundational concepts, practical tips, and industry trends from the current "golden era" of AI. Sam walks through core chapters of his guide, demystifying agents, prompt engineering, tool use, workflows, RAG (retrieval-augmented generation), and the rapid progress being made by startups across industries.
- Building AI agents now feels like the early days of the internet (circa 1996)
- Surging interest and new startups across every vertical—AI is relevant everywhere
- Founders are discovering fundamental, sometimes counterintuitive, AI engineering principles
- Break complex tasks into simpler LLM calls and connect them with workflow graphs
- Use more examples in prompts—few-shot learning boosts output quality dramatically
- Start simple: use OpenAI for most tasks, Claude for code, and avoid overcomplicating with infrastructure early
- Agents operate on a spectrum from simple function callers to fully autonomous systems with memory and task creation
- Too much autonomy can harm accuracy; sometimes it's better to restrict agent scope
- Structured workflow graphs help coordinate multi-agent tasks and clarify responsibilities
- Define tools (functions) with clear, descriptive names and provide example uses
- System prompts should include detailed tool descriptions for better agent performance
- Multi-agent systems can use supervisor models or hand-off via structured workflow graphs for coordination
- RAG enables semantic search and retrieval, not just keyword matching
- The process involves chunking, embedding, vector storage, retrieval, and reranking
- Debate exists on RAG's future as LLM context windows expand, but it's still a critical skill today
- Startups are applying agents to diverse fields: aerospace, medical transcription, finance, and more
- Foundational skills: prompt design, memory management, tool integration, and workflow structuring
- "Principles of Building AI Agents" is positioned as a practical guide for new and experienced builders