A researcher's firsthand account of the ACM CAIS (Conference on AI and Agentic Systems) in San Jose, covering key themes across three days of talks and workshops. The conference centered on compound AI architectures, multi-agent coordination, and operational concerns like evaluation, cost optimization, and governance. Workshop highlights include OpenHands' skill induction approach, RL environment lifecycle design, and keynotes on agents as database users (ClickHouse), automated database tuning vs. LLM-based tuning (Andy Pavlo), and serverless infrastructure for agents (Neon/Databricks). A notable paper deep-dive covers TraceFix, which uses TLA+ formal verification to eliminate deadlocks and race conditions in multi-agent workflows, achieving 100% verification success and dramatically faster execution than unstructured chat-based coordination. The author critiques the conference's overly broad scope and the industry trend toward ever-shorter presentation slots.
