Trustworthy, Safe, Efficient and Reliable Agents

The Data, Agents, and Processes Lab (DAPLab) at Columbia University develops systems, infrastructure, and interaction principles for AI agents to safely and reliably automate real work.

Agents turn every layer of the stack into a semantic quality problem. As a consequence, trustworthy agent automation cannot be solved at a single layer of the stack. DAPLab vertically integrates expertise across operating systems, data systems, AI, HCI, security, and enterprise workflows to rethink what agent development, evaluation, and interaction, along with the computing infrastructure should look like. We work closely with industry partners to ground this research in real organizational needs and ensure it delivers practical impact.

For more information about the lab, contact ewu@cs.columbia.edu.

Why vertical integration matters

AI agents fail across the entire stack: models hallucinate, retrieval misses critical context, execution environments lack isolation, workflows leak data, and human oversight breaks under scale. Fixing only one layer is not enough.

DAPLab brings together researchers across systems, databases, AI, HCI, security, and organizational workflows because trustworthy automation requires coordinated advances across the full agent stack — from infrastructure and state management to evaluation, safety, and human interaction.

  1. Human Interaction
  2. Agents + Models
  3. Processes + Assurance
  4. Data + Execution Systems
  5. OS + Infrastructure

News

All news →

  • Asaf Cidon and coauthors win SOSP 2026 Best Paper and Distinguished Artifact

    The paper “It’s the Kernel’s Fault! Custom Page Fault Handling With bpf_fault” by Tal Zussman, Riju Dey, Hasan Zengin, Yiming Fang, David Hildenbrand, and Asaf Cidon received the SOSP 2026 Best Paper and Distinguished Artifact awards.

  • SceniX acquired by World Labs; AMD to acquire World Labs

    Huge congrats to DAPLab faculty Yunzhu Li and Changxi Zheng, whose robotics startup, SceniX, was acquired by World Labs in July 2026. In September a few months later, AMD announced an agreement to acquire World Labs, bringing the SceniX team’s work on robotics, simulation, and spatial intelligence into AMD.

  • Fall 2026 DAPLab Research Seminar

    The DAPLab’s Tuesday 12PM research seminar continues in CSB 453 (CS Conference Room). We invite internal and external speakers that can share cutting-edge agent-systems research or can talk about processes in their organizations and how they are trying to automate them.

  • MortarBench featured in Realtor.com

    Realtor.com covered our MortarBench research on AI mortgage origination agents. Top models got nearly 1 in 4 answers wrong under realistic conditions — and showed systematic bias, flagging non-English names as “foreign origin” at 5× the rate of English names. Matthew Toles and Zhou Yu are quoted in the piece.

  • Haonan Wang receives inaugural Workday AI PhD Fellowship

    Haonan Wang has been awarded the inaugural Workday AI PhD Fellowship to pursue his work on Data Lake Agents. The core question: how do you teach an enterprise agent to gather sufficient, traceable evidence from an organization’s heterogeneous data sources and correctly execute the resulting workflow — while satisfying regulatory requirements, organizational policies, role-based permissions, approval hierarchies, fairness constraints, and operational budgets?

Events

All events →

Upcoming

Recent

  • Sep 29, 2026

    OpenAgents: The Next-Gen Collaboration OS for Teams of Humans and Agents

    Raphael Shu

  • Sep 24, 2026

    Selective interventions for information acquisition: Incentives and audits for biodiversity

    Lily Xu

  • Sep 15, 2026

    Columbia AI Night

  • Sep 4, 2026

    Fall 2026 Personal Health Assistant Course

    Xuhai "Orson" Xu, Eugene Wu

Publications

All →

Thanks to our industry partners