Agents in the Wild

abstract.md

$ cat abstract.md

Agentic systems — LLM-based architectures capable of reasoning, planning, acting, and coordinating with tools and other agents — are rapidly transitioning from research prototypes to production-scale deployments across software engineering, scientific discovery, and finance.

This tutorial brings together researchers and practitioners to explore advances in reasoning and planning, multi-agent coordination, and evaluation — with a sharp focus on the messy, hard problems that only appear once agents leave the lab and enter the wild.

Through case studies in pharmaceutical discovery and financial systems, we analyze common design patterns, and discuss practical mitigation strategies for failure modes: verification pipelines, fallback mechanisms, and human-in-the-loop supervision.

$▋

What you'll learn

Five Things You'll Take Home

The full arc — from single-LLM prompting to modular multi-agent orchestration — and why the shift happened.

Modern reasoning, planning, and execution strategies with honest trade-off comparisons.

Evaluation beyond benchmarks — robustness, reliability, safety, and continuous adaptation in production.

Concrete deployment case studies — hallucination, deadlocks, drift, cascading errors — and how teams survived them.

Open research frontiers at the intersection of academic innovation and industrial reality.

Schedule

Tutorial Program

Part I — Foundations

3.1

Agentic Systems: History & Definitions

From single-model pipelines combining prompting, tool use, and heuristic control — to modular, multi-agent architectures where reasoning, planning, execution, and memory interlock.

3.2

Reasoning, Planning & Multi-Agent Coordination

Task decomposition, multi-plan generation, iterative reflection, memory-augmented planning. Architectural topologies — peer-to-peer, hierarchical — communication protocols, and fault-tolerance for agent teams.

3.3

Retrieval and Reasoning Pipelines

Iterative, adaptive, and modular retrieval strategies that interleave reasoning and information lookup for multi-step, dynamic, open-ended tasks.

3.4

Evaluation Beyond Benchmarks

Dynamic, behavior-centric frameworks. Scenario-driven robustness, behavioral safety, trustworthiness. Concrete analyses of hallucination, deadlocks, drift, and cascading errors — with mitigation playbooks.

Part II — Applied Perspectives

4.1

Agents in Pharmaceutical & Life Sciences

End-to-end autonomous scientific discovery spanning literature synthesis, hypothesis generation, and experimental validation. Multi-agent frameworks for drug candidate identification and biomedical discovery.

4.2

Agents in Finance

From static analytical tools to dynamic decision-support systems. Planner–executor–verifier architectures for earnings call summarization, portfolio allocation, and risk assessment.

The team

Organizers

Grace Hui Yang

Georgetown University Washington, D.C., USA

Professor of Computer Science leading the InfoSense research group. NSF CAREER Award recipient. Research bridges information retrieval and NLP — conversational AI agents, retrieval-augmented generation, and deep RL for dialogue. General Co-Chair of SIGIR 2024.

grace.yang [at] georgetown.edu

Pranav N. Venkit

Salesforce San Francisco, California, USA

Research Scientist at Salesforce AI Research, building trustworthy interactive agents with long-term reasoning and memory. Ph.D. in Informatics from Penn State (Best AI Dissertation). Research spans NLP, HCI, social informatics, and privacy.

pnarayananvenkit [at] salesforce.com

Hooman Sedghamiz

Bayer San Diego, California, USA

Sr. Director of AI/ML at Bayer AG, leading enterprise-scale generative AI and agentic systems for precision medicine (100,000+ employees). Co-Chair EMNLP 2023 GEM Industrial Track. Published in EMNLP, BMC Systems Biology, Frontiers. International patent holder (BioSigKit).

hooman.sedghamiz [at] bayer.com

Enrico Santus

Bloomberg New York City, New York, USA

Principal Technical Strategist for QUANT NLP in Bloomberg's Office of the CTO. 17+ years building AI products across medicine, finance, and NLP. Ph.D. in Computational Linguistics (HKPolyU); postdoc at SUTD & MIT. Invited to speak on NLP at the White House; co-authored AI/ML factsheets for the US Congress.

esantus [at] bloomberg.net

Victor Dibia

Microsoft Research Seattle, Washington, USA

Principal Research Software Engineer at Microsoft Research / Core AI. Creator of Magentic-UI, Magentic-One, AutoGen Studio, and LIDA (ACL 2023). Ph.D. in Information Systems from City University of Hong Kong; M.S. from Carnegie Mellon University.

victordibia [at] microsoft.com

Ioana Baldini

Bloomberg New York City, New York, USA

Research Scientist on the AI Strategy & Research team in Bloomberg's Office of the CTO. Former IBM Research career spanning NLP, cloud infrastructure, and heterogeneous computing. NSERC Graduate Scholarship, IBM Ph.D. Fellowship, Canada Google Anita Borg Scholarship. Ph.D. from the University of Toronto.

ibaldinisoar [at] bloomberg.net

Agentic Systems Large Language Models Multi-Agent Coordination Autonomous Workflows Agent Evaluation Robustness & Safety Pharmaceutical Discovery Financial AI Human-in-the-Loop Reasoning & Planning