AgentCraft: Workshop on Developing Trustworthy Agentic AI Systems

Ambitious efforts are underway to build AI agents powered by large language models (LLMs) across a spectrum of domains, including programming, productivity, creative work, and enterprise automation. Although many frameworks have emerged to aid in the agent development process, major challenges remain regarding agent autonomy, reasoning processes, unpredictable behaviors, and consequential actions. Developers struggle to comprehend and debug their agent’s behavior in systematic ways. They also struggle to determine where in an agent’s workflow human input and oversight are needed. Intelligent interfaces that enable meaningful oversight of agentic plans, decisions, and actions are needed to foster transparency, build trust, and manage complexity. Further, evaluations of agentic AI systems require moving past automated machine learning benchmarks toward human-centric evaluations that demonstrate real-world usefulness, reliability, and safety.