Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards Enterprise-Ready Computer Using Generalist Agent

Sami Marreed, Alon Oved, Avi Yaeli, Segev Shlomov, Ido Levy, Offer Akrabi, Aviad Sela, Asaf Adi, Nir Mashkif

TL;DR

The paper presents IBM's Computer Using Generalist Agent (CUGA), an enterprise-focused, multi-agent system designed to operate across web and API tasks with emphasis on privacy, safety, and cost-efficiency. It details an iterative, failure-driven development approach that combines Plan Controller and Sub-task Plan-Execute agents, advanced grounding, and knowledge-injection techniques to tackle long-horizon tasks. Achieving state-of-the-art results on WebArena and AppWorld, the work demonstrates significant progress in stability, scalability, and generalization, while acknowledging challenges in safety, benchmarking realism, and enterprise deployment. The authors also outline a practical roadmap for advancing evaluation frameworks, automation of failure analysis, and deployment in real-world enterprise environments.

Abstract

This paper presents our ongoing work toward developing an enterprise-ready Computer Using Generalist Agent (CUGA) system. Our research highlights the evolutionary nature of building agentic systems suitable for enterprise environments. By integrating state-of-the-art agentic AI techniques with a systematic approach to iterative evaluation, analysis, and refinement, we have achieved rapid and cost-effective performance gains, notably reaching a new state-of-the-art performance on the WebArena and AppWorld benchmarks. We detail our development roadmap, the methodology and tools that facilitated rapid learning from failures and continuous system refinement, and discuss key lessons learned and future challenges for enterprise adoption.

Towards Enterprise-Ready Computer Using Generalist Agent

TL;DR

The paper presents IBM's Computer Using Generalist Agent (CUGA), an enterprise-focused, multi-agent system designed to operate across web and API tasks with emphasis on privacy, safety, and cost-efficiency. It details an iterative, failure-driven development approach that combines Plan Controller and Sub-task Plan-Execute agents, advanced grounding, and knowledge-injection techniques to tackle long-horizon tasks. Achieving state-of-the-art results on WebArena and AppWorld, the work demonstrates significant progress in stability, scalability, and generalization, while acknowledging challenges in safety, benchmarking realism, and enterprise deployment. The authors also outline a practical roadmap for advancing evaluation frameworks, automation of failure analysis, and deployment in real-world enterprise environments.

Abstract

This paper presents our ongoing work toward developing an enterprise-ready Computer Using Generalist Agent (CUGA) system. Our research highlights the evolutionary nature of building agentic systems suitable for enterprise environments. By integrating state-of-the-art agentic AI techniques with a systematic approach to iterative evaluation, analysis, and refinement, we have achieved rapid and cost-effective performance gains, notably reaching a new state-of-the-art performance on the WebArena and AppWorld benchmarks. We detail our development roadmap, the methodology and tools that facilitated rapid learning from failures and continuous system refinement, and discuss key lessons learned and future challenges for enterprise adoption.

Paper Structure

This paper contains 18 sections, 6 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Methodology and Tools
  3. Architecture Evolution
  4. Addressing Long-Horizon and Complex Tasks
  5. Sub-Execution Agents Challenges and Enhancements
  6. Browser Sub-task Agent Challenges
  7. Enhancements for Browser Sub-task Agent
  8. API Sub-task Agent Challenges
  9. Enhancements for API Sub-task Agent
  10. Enhancing Stability and Mitigating Hallucinations
  11. Planner Alignment through Context Enrichment, Learning, and Knowledge Injection
  12. Results
  13. Lessons Learned, Challenges, and Roadmap
  14. Methodology and Tools:
  15. Evaluating Generalist Agents on Realistic Benchmarks:
  16. ...and 3 more sections

Figures (6)

  • Figure 1: CUGA’s iterative evaluate-analyze-enhance cycle, enabling rapid failure diagnosis, targeted improvements, and continuous performance gains.
  • Figure 2: CUGA performance dashboard providing an overview and detailed performance results per task, with drill-down links into trajectories
  • Figure 3: The CUGA trajectory visualization on the AppWorld task.
  • Figure 5: A simplified high-level representation of IBM CUGA architecture, illustrating the interaction between user intent, context enrichment, high-level plan controller, sub-task plan-execute agents, environment action and observation, and learning and knowledge components.
  • Figure 6: Architecture of the API and Browser Sub Agents
  • ...and 1 more figures