From Benchmarks to Business Impact: Deploying IBM Generalist Agent in Enterprise Production

TL;DR

The paper investigates how to translate generalist computer-using agents from benchmark success to enterprise production value. It presents the CUGA architecture, a layered planner–executor system, and reports state-of-the-art results on AppWorld and WebArena, plus a pilot in BPO Talent Acquisition using the BPO-TA benchmark. It highlights enterprise requirements—safety, efficiency, integration, and governance—and distills technical and organizational lessons from Phase 1 to guide enterprise deployment. Preliminary results indicate substantial potential in reducing development effort and improving auditability, while maintaining accuracy near specialized agents. The work charts a path for adoption of enterprise-ready generalist agents and identifies future steps for production-grade architectures.

Abstract

Agents are rapidly advancing in automating digital work, but enterprises face a harder challenge: moving beyond prototypes to deployed systems that deliver measurable business value. This path is complicated by fragmented frameworks, slow development, and the absence of standardized evaluation practices. Generalist agents have emerged as a promising direction, excelling on academic benchmarks and offering flexibility across task types, applications, and modalities. Yet, evidence of their use in production enterprise settings remains limited. This paper reports IBM's experience developing and piloting the Computer Using Generalist Agent (CUGA), which has been open-sourced for the community (https://github.com/cuga-project/cuga-agent). CUGA adopts a hierarchical planner--executor architecture with strong analytical foundations, achieving state-of-the-art performance on AppWorld and WebArena. Beyond benchmarks, it was evaluated in a pilot within the Business-Process-Outsourcing talent acquisition domain, addressing enterprise requirements for scalability, auditability, safety, and governance. To support assessment, we introduce BPO-TA, a 26-task benchmark spanning 13 analytics endpoints. In preliminary evaluations, CUGA approached the accuracy of specialized agents while indicating potential for reducing development time and cost. Our contribution is twofold: presenting early evidence of generalist agents operating at enterprise scale, and distilling technical and organizational lessons from this initial pilot. We outline requirements and next steps for advancing research-grade architectures like CUGA into robust, enterprise-ready systems.

Figures (7)

• Figure 1: Reducing time-to-value with generalist agents. Traditional specialized agents (left) require extensive custom design and benchmarking. Generalist agents (right), benchmarked on complex tasks such as AppWorld, WebArena, and TauBench2, shift the enterprise focus to configuration and domain-specific evaluation.
• Figure 2: CUGA general architecture via Layered planner–executor loops.
• Figure 3: Architecture of the Browser Sub Agent: the Browser Planner combines action history and reflection to select between two execution paths--an Action Agent (click, type, select, navigate) and a Question Answering Agent (DOM-to-Markdown conversion and screenshots).
• Figure 4: Architecture of the API Sub Agent: the API Planner coordinates short-term memory, reflection, and two executors--the ShortlisterAgent, which routes to APIs through a registry, and the CodeAgent, which generates and executes code via a nested CodePlanner.
• Figure 5: WebArena intent distribution across different websites. Cross-site intents include interaction with multiple websites.