Build real agentic apps using CUGA: two dozen working examples on a lightweight harness

2026-06-23 · Source: Hugging Face - Blog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Intermediate, long

Summary

IBM has released CUGA (Configurable Generalist Agent), an open-source agent harness designed for enterprise applications, which simplifies building agentic apps by handling orchestration, planning, execution loops, and state management. CUGA allows developers to focus on defining tools and prompts, rather than plumbing. To demonstrate its capabilities, IBM developed cuga-apps, a collection of two dozen single-file agent applications, ranging from a movie recommender to an IBM Cloud architecture advisor, available in a live gallery. CUGA has topped agent benchmarks like AppWorld (07/25 - 02/26) and WebArena (02/25 - 09/25) by performing planning, reflection, and variable-tracking itself, enabling smaller open-weight models like gpt-oss-120b to perform effectively. The harness also integrates a policy system for governance, including Intent Guards and Tool Approvals, and supports multi-agent delegation via "CugaSupervisor" and "Agent Skills" for complex tasks. This architecture facilitates seamless deployment from development to governed production environments like IBM Sovereign Core, where agents run in transient, isolated containers with in-tenant telemetry.

Key takeaway

For AI Engineers building agentic applications, CUGA offers a streamlined approach to development and production deployment. You can rapidly prototype by defining only tools and prompts, avoiding extensive plumbing. This harness integrates governance and multi-agent capabilities by design, ensuring your agents are production-ready and secure without requiring a rewrite. Consider adopting CUGA to accelerate your agent development lifecycle and ensure compliance from the outset.

Key insights

CUGA simplifies agent development by abstracting orchestration, enabling focus on tools and prompts for robust, governed applications.

Principles

• Agent orchestration should be handled by a configurable harness.
• Explicit failure handling prevents agent run derailment.
• Governance policies should be integrated into the agent runtime.

Method

Build a "CugaAgent" by defining a tool list and a system prompt, then "await agent.invoke(...)". For complex tasks, use "CugaSupervisor" to delegate to specialist agents.

In practice

• Use "pip install cuga" to quickly set up an agent development environment.
• Clone "cuga-apps" to access two dozen working agent examples.
• Implement Intent Guards to block unwanted agent actions.

Topics

• Agentic AI
• CUGA
• IBM Sovereign Core
• Multi-agent Systems
• LLM Governance
• FastAPI

Code references

• cuga-project/cuga-agent
• cuga-project/cuga-apps
• cuga-project/cuga-apps

Best for: AI Architect, CTO, VP of Engineering/Data, AI Engineer, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• The Emerging "Harness Engineering" Playbook — Effective AI agent integration requires dedicated "harness engineering" to structure environments and manage agent workflows.
• Introducing Kasal — Kasal democratizes agentic AI development through a visual, Databricks-native platform for diverse skill levels.
• Add a Specialized Deep Research Skill to Agent Harnesses — NVIDIA AI-Q provides agent harnesses a dedicated, secure, and auditable deep research skill for enterprise data synthesis and analysis.
• Whitepaper Companion Podcast - Prototype to Production — Productionizing AI agents requires specialized AgentOps, focusing on infrastructure, security, and continuous validation beyond core AI development.
• Day 5 - Agentic Engineering/VIbe Coding an AI App from Zero To Production | 🔴 Live — AI agents can achieve high test coverage and implement features in complex applications.
• How Harness-as-a-Service Will Change Agents — The AI boom is evident in big tech earnings, while "harness as a service" is democratizing agent development by abstracting runtime complexities.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Hugging Face - Blog.