CUGA: A Lightweight Harness for Enterprise Agent Apps
Use CUGA to build an agent by defining a tool list and prompt in a single FastAPI file, leveraging its built‑in planning and state management.
Clone the cuga-apps repo, pick an example, and run it with `pip install cuga` and your chosen LLM provider.
Summary
On June 23 2026, IBM announced CUGA (Configurable Generalist Agent), an open‑source agent harness that abstracts planning, execution, and state management, allowing developers to write a tool list and prompt in a single file. The library ships with a FastAPI wrapper and supports multiple LLM providers (OpenAI, Anthropic, watsonx, Ollama, LiteLLM) via an environment‑variable‑driven factory. CUGA’s core features include long‑horizon planning, reflection, variable tracking, and code execution in sandboxed environments, enabling it to outperform benchmark agents like AppWorld and WebArena. IBM released a gallery of 24 lightweight apps, each a single FastAPI file, ranging from a movie recommender to an IBM Cloud architecture advisor, all using the same CUGAAgent constructor. The harness handles tool calls, guardrails, and multi‑agent delegation, while developers only need to define inline tools and prompts. The IBM Cloud advisor example demonstrates a tool that queries the MCP catalog API, ensuring recommendations are valid. CUGA also supports declarative guardrails, different reasoning modes (Fast, Balanced, Accurate), and can run locally or in the cloud. The release positions CUGA as a production‑ready agent framework that reduces boilerplate and accelerates agentic app development.
Key changes
- Release of CUGA (Configurable Generalist Agent) harness for agent development.
- Supports multiple LLM providers via environment‑variable factory.
- Core features: long‑horizon planning, reflection, variable tracking, sandboxed code execution.
- 24 single‑file FastAPI apps released, including IBM Cloud advisor and movie recommender.
- Tool calls handled by CUGA; developers only define inline tools and prompts.
- Declarative guardrails and reasoning modes (Fast, Balanced, Accurate) configurable.
- Supports multi‑agent delegation and MCP tool integration.
- Positions CUGA as a production‑ready, low‑boilerplate agent framework.