Expert training and strategic guidance for responsible AI adoption

Even with a strong grasp of how LLMs work, using them well requires a great deal of effort and intention. Getting reliable output means carefully structuring many prompts, providing the right context and files, checking results, and then repeating the whole process again for each task -- and small errors compound quickly across multi-step work. This workshop introduces attendees to the more scalable approach of "agent orchestration." Instead of directly overseeing all tasks in one sustained conversation with an assistant, attendees will learn how to leverage an "orchestrator" agent that modularizes and distributes tasks across multiple parallel assistants ("subagents") on their behalf. Through hands-on exercises, attendees will practice launching, supervising, and critically reviewing this style of multi-agent workflow themselves -- using the Data Analyst Augmentation Framework (DAAF), an open-source research orchestration framework built on Claude Code, as a concrete exemplar of these practices. Throughout, the workshop helps attendees navigate the tension between greater automation and auditability while firmly staying in control of the decisions that matter most.

Once attendees can leverage AI orchestration frameworks, the core question of practice shifts from "how do I do this task?" to "how do I do this task most effectively and efficiently?" This workshop teaches the deeper mechanics that answer that question. Attendees will learn to see their AI systems as three distinct and interacting layers -- the model itself, the "harness" that runs it (like Claude Code), and the custom frameworks they build on top -- and how understanding each layer unlocks more capable, more reliable workflow designs. From there, we cover "hooks" (small programs that enforce your rules automatically, rather than just asking the AI to follow them) and "architectural prompting" (using AI to help design and improve the instruction systems that govern its own behavior). A real-world case study then shows how these techniques combine to creatively work around even hard model limitations. The session closes with a design clinic applying these expanded design principles to attendees' own problems -- moving attendees from users of these frameworks to confident architects of systems they can trust.

AI agents can now autonomously plan, write, review, and execute analytic code, raising urgent questions about their role in research given known risks like hallucinations and inaccuracies. This session provides researchers with an approachable framework for realizing the benefits of AI agents for more complex research workflows, while mitigating their very real risks, using DAAF: an open-source set of tools for Claude Code designed to enhance and support data analysis while strictly enforcing transparency, reproducibility, and rigor. In addition to building familiarity with how to use and extend DAAF for their own purposes, attendees will leave with a strong understanding of the pros and cons of current AI agent frameworks for data analysis workflows, the main design considerations involved in robust "context engineering" practices, what these tools currently can and cannot do, as well as the intuition needed to evaluate their appropriateness for their own work.

Just as AI tools are rapidly reshaping what it means to practice data science and statistics, so too are they reshaping what it means to teach data science and statistics. Professors and instructors are now faced with an extremely difficult challenge: How can you redesign your curriculum to respond to the expanding role of AI in students' lives without knowing what kinds of jobs and careers will exist for them in the future? What skills and topics and experiences will be valuable for students beyond just the next few months? This workshop brings faculty together to work through those design tensions directly. We'll introduce practical heuristics for distinguishing "durable" skills in the era of AI, explore targeted opportunities to modernize curricula and assignments without abandoning the conceptual depth that makes graduates genuinely capable, and develop a more informed critical lens for evaluating AI developments as they continue to emerge. Participants come having pre-identified core "problems of practice" from reviewing their syllabi for the upcoming semester, so that the session's facilitated discussions center on real design tensions rather than hypotheticals -- questions like what trade-offs exist between productive struggle and AI assistance in the classroom, what role assessment should play when students can generate plausible-looking work effortlessly, and what opportunities AI presents for richer course materials and activities. Attendees will leave with concrete strategies for more agile curriculum adaptation and a stronger framework for making principled instructional decisions as the landscape continues to shift.