IBM's CUGA harness lets you build agentic apps with just a tool list and prompt

IBM Research released CUGA (Configurable Generalist Agent), an open-source agent harness that inverts the typical agent-building workflow. Instead of spending a week on wiring model clients, tool adapters, state streaming, and orchestration, developers only need to define a tool list and a prompt. The harness handles planning, execution loops, tool calls (OpenAPI, MCP, LangChain functions), variable tracking across long tasks, and a reflection step that catches bad calls and re-plans. This architecture is why CUGA topped the AppWorld leaderboard from July 2025 to February 2026 and WebArena from February to September 2025. Developers can choose from Fast, Balanced, and Accurate reasoning modes via config, and run code in local, Docker/Podman, or E2B cloud sandboxes. The same agent definition works across OpenAI, watsonx, Ollama, and other providers by changing a single environment variable.

To prove its simplicity, IBM published cuga-apps—a gallery of two dozen working single-file FastAPI applications, from a movie recommender to an IBM Cloud architecture advisor. Each app wraps one CugaAgent instance with a tool list and prompt, making them easy to read, copy, and adapt. CUGA also includes declarative guardrails, multi-agent delegation over A2A, and Docling-powered RAG. The key insight: rather than relying on frontier models to recover from planning errors, CUGA's harness carries that load itself, enabling smaller open-weight models like gpt-oss-120b to perform competitively. For enterprises, this means sovereign, governed agent deployment without rewriting the agent logic.