Beyond Prompts: Building Interruptible, Cooperative AI Agents in Production

As AI adoption accelerates, the focus is shifting from single-shot prompts and chat interfaces to persistent, intelligent agents that can plan, act, and collaborate with both humans and tools. But deploying such agents in production—especially in infrastructure or DevOps-heavy contexts—requires more than LLM magic. It requires architecture.

This talk walks you through how to build interruptible, tool-calling AI agents that can reason over tasks, handle human input in the loop, cooperate with other agents, and call external tools or APIs dynamically—using modern techniques like Model Context Protocol (MCP), loop debugging, and event-based interruptions.

We’ll break down:

• Agent loop architectures (e.g. plan-act-reflect)
• Tool calling: how agents decide when and how to call real infrastructure APIs
• Interruptions and reversibility: critical for trust and user control
• Agent-human vs agent-agent collaboration: use cases and failure points
• Debugging agent workflows via MCP: trace, pause, rerun

Using examples from real-world platforms (like AI-powered DevOps assistants), this session provides mental models and code patterns to make your agents reliable, debuggable, and aligned with real-world user expectations.