Building Responsible Autonomous AI Agents: Principles, Patterns, and Practices
summary
Shows a concrete agentic system called BioTracker that ties Fitbit/Wyze data into a Claude-backed chat API, a 12-tool MCP server, and a GitHub Copilot SDK reporting pipeline that generates Python visualizations and PDF reports. Demonstrates Entra Agent ID for mutual agent authentication, OpenTelemetry logging for every tool call, middleware enforcing tool budgeting and whitelists, reviewer sub-agents, kill-switch patterns, PyRITE red‑teaming and Azure AI Content Safety for groundedness and prompt‑injection defenses.
AI-written summary · about
chapters (13)
- 0:00Responsible AI agents intro
- 3:57Risks in autonomous agents
- 11:48Microsoft's AI principles
- 13:16Agentic systems tech stack
- 21:23Review agents and fail-safes
- 24:54Defense-in-depth security
- 34:51Tool call security
- 37:20Fairness testing and guardrails
- 42:13System prompt guardrails
- 45:01Observability and monitoring
- 46:45Red teaming and evaluation
- 48:14AI regulations and compliance
- 49:40Key takeaways
related talks
40mBuilding Effective AI Agents Simple, Composable Patterns
40mAI Agents in Practice
22mHOW TO BUILD YOUR OWN AI AGENT: THE COMPLETE ARCHITECTURE FOR ENGINEERING TEAMS
30mBuilding Reliable AI Agents: An Open Source Approach To Evaluation and Observability
01h 04mBuilding AI Applications the Pydantic Way
40mAgents in Action: Advancing Open Source AI Missions
44mBuilding an AI Agent on Top of Your API
33mWhat Did My Agent Do? Observability and Accountability for AI Agents