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

Source: HuggingFace Date: 2026-06-23 URL: https://huggingface.co/blog/ibm-research/cuga-apps

Summary

IBM Research published CUGA (Configurable Generalist Agent), an open-source agent harness that abstracts planning, execution loops, tool dispatch, and state management so developers can define tools and prompts without rebuilding the infrastructure layer. The cuga-apps gallery ships two dozen single-file FastAPI applications — movie recommender, cloud advisor, research scouts, a seven-agent lead-gen pipeline — all designed to be read, copied, and adapted. Provider switching (OpenAI, Anthropic, watsonx, Ollama, etc.) is one-line; the harness includes a declarative policy system with six guardrail types and multi-agent delegation.

Implications

• Agent landscape: CUGA sits in the same tier as LangGraph, CrewAI, and AutoGen — a coordination layer that lets you define agent topology without writing the loop yourself. IBM publishing it as open-source with a live app gallery is a substantive attempt to shape harness conventions, not just a research artifact.
• Dev-tooling floor: Single-file FastAPI patterns with a minimal agent API lower the floor for teams that want agentic behavior without adopting a heavy framework. The gallery-first release strategy (clone and run) is a direct counter to documentation-heavy SDK approaches.
• Voices/power dynamics: IBM Research shipping an Anthropic-compatible harness while also listing watsonx as a provider is a positioning move: claim the open standards table while keeping enterprise customers on the IBM stack. Watch whether the open-source community adopts CUGA independently or treats it as IBM-flavored.