12-Factor Agents: Patterns of reliable LLM applications — Dex Horthy, HumanLayer

The speaker discusses the challenges and patterns observed while building AI agents, emphasizing that many production agents are not very "agentic" and that certain core patterns make LLM-based applications effective. The speaker introduces the concept of "12 Factors of AI Agents," a set of guidelines derived from interviews with numerous founders, builders, and engineers. The talk advocates for rethinking agent development from first principles, applying software engineering practices to build reliable agents, and focusing on key factors such as JSON parsing, controlling prompts and context windows, managing control flow, and integrating human input. The speaker also touches on the use of microagents within deterministic workflows and encourages developers to find the boundaries of what models can reliably do to engineer magical and superior systems.