Source: arXiv (KDD 2026 tutorial) — 2026-07-21
Summary
"Agents in the Wild" is a tutorial-survey, presented at the 32nd ACM SIGKDD Conference (KDD 2026), authored by researchers from Georgetown University, Salesforce, Bayer, Bloomberg, and Microsoft Research. It surveys how agentic LLM systems — architectures that reason, plan, act, and coordinate with tools and other agents — move from research prototypes to production deployments across software engineering, scientific discovery, and finance, arguing academic benchmarks miss the robustness and reliability failures that only surface once agents run at production scale.
Key Takeaways
- Frames a real gap: academic agent research optimizes for benchmark scores, while production deployment surfaces different failure classes entirely — robustness, safety, and reliability under real-world input distributions.
- Covers applied case studies in pharmaceutical discovery (Bayer) and financial systems (Bloomberg), grounding the survey in named industry deployments rather than only synthetic benchmarks.
- Surveys practical mitigation patterns teams are actually using: verification pipelines, fallback mechanisms, and human-in-the-loop supervision as the standard toolkit for hardening agents post-lab.
- Positioned as a tutorial (not a novel method paper) — useful as a map of the current deployment landscape and failure taxonomy, not a new technique to adopt directly.