IBM has open sourced CUGA, an enterprise-grade AI agent that completes more than half of complex business tasks.
IBM has released CUGA (Configurable Generalist Agent) as an open-source AI agent, aimed at automating complex enterprise workflows through multi-agent orchestration, API integration, and code generation. Licensed under Apache 2.0, CUGA can be freely used, modified, and deployed in enterprise environments without restrictive licensing barriers.
Benchmark results position CUGA among the most capable AI agents currently available. The agent achieved a 61.7 percent success rate on the WebArena benchmark for web-based tasks and a 48.2 percent scenario completion rate on AppWorld, which evaluates API-driven workflows. IBM researchers describe these figures as top-tier performance for AI agents, even though they remain far from human reliability.
CUGA is built specifically for enterprise use, featuring intent detection through a chat layer, structured task planning, a dynamic task ledger that enables re-planning, delegation to specialised agents, secure sandboxed code execution, and policy-compliant response generation. The system integrates with Langflow, a low-code agent design platform, and supports open models including gpt-oss-120b and Llama-4-Maverick-17B-128E-Instruct-fp8.
In a research paper, IBM authors stated: “Our vision for IBM CUGA is to develop a generalist agent that can be adapted and configured by knowledge workers to perform routine or complex aspects of their work in a safe and trustworthy manner.”
IBM acknowledges broader industry challenges. Its own WebAgentBench research shows that enterprise agents average 24.4 per cent raw task completion and just 15 per cent policy-compliant completion, dropping to 7.1 per cent when multiple policies apply. Despite known limitations, including a reported run-loop bug, IBM argues that open sourcing CUGA prioritises real-world measurability, transparency, and collaborative progress over closed, opaque agent systems.
