Building Applications with AI Agents
summary
Generative models can reach ~70% out-of-the-box on complex agentic tasks but climbing past that long tail requires architecture and evaluation: register few tools per prompt because empirically more tools reduces accuracy, group APIs into human‑facing actions, prefer chains/trees before full agentic loops, and split functionality into semantically similar agents with a coordinator for scale. Invest in evaluation and observability using Label Studio, Intel Agent for synthetic cases, Microsoft’s Pirate for red‑teaming, OpenTelemetry tracing, and automated failure clustering — all while building tripwires for security.
AI-written summary · about
chapters (12)
- 0:00Introduction by Michael Albada, Principal Applied Scientist at Microsoft.
- 1:14The Promise and Obstacles of Agentic Development.
- 2:37Defining What an AI Agent Is (and Isn't).
- 4:42Core Component 1: Tool Use and Function Calling.
- 6:37Core Component 2: Orchestration Patterns (Chains, Trees, Agentic).
- 8:47Core Component 3: Multi-Agent Systems.
- 9:43Common Pitfall 1: Insufficient Evaluation.
- 11:25Overview of specific Evaluation Tools.
- 12:57Common Pitfall 2: Lack of Observability.
- 13:50Other Common Pitfalls (Tool issues, complexity).
- 14:45The Critical Importance of Security and Safety.
- 15:15Conclusion and Future Outlook.
related talks
28mAgentic AI
31mTesting, Evaluating and Debugging Generative AI and Agentic Ap...
26mAI Agents and the Future of Human-AI Collaborative Intelligence
32mApproach to Architecting Generative AI Solutions
22mHOW TO BUILD YOUR OWN AI AGENT: THE COMPLETE ARCHITECTURE FOR ENGINEERING TEAMS
29mBuilding Trust in AI Agents
21mRevolutionizing software creation: Harnessing the power of GenAI
44mBuilding Agentic AI Systems: From Transactions To Conversations