IBM open-sources CUGA agent harness

What happened

IBM open-sourced CUGA (Configurable Generalist Agent), a lightweight harness that handles planning, execution loops, tool calls, and state management for AI agents. The company built two dozen working example apps—each a single file—to show how little code developers need to write once the harness takes care of the scaffolding.

Why it matters

Most agentic apps require a week of plumbing before they do anything useful; CUGA inverts that by pre-assembling the complex parts (long-horizon planning with variable tracking, self-correction reflection steps, guardrails, multi-agent delegation) so developers focus only on defining tools and writing the prompt. The harness has topped agent benchmarks like AppWorld and WebArena, and lets smaller open-weight models like gpt-oss-120b work where frontier APIs would normally be required.

What to watch

CUGA accepts tools via OpenAPI, MCP (Model Context Protocol), and LangChain functions; runs on OpenAI, Anthropic, watsonx, LiteLLM, or Ollama depending on an environment variable; and offers three reasoning modes (Fast, Balanced, Accurate) so the same agent definition can dial cost and latency without a rewrite. It is available via pip install cuga.