Build real agentic apps using CUGA: two dozen working examples on a lightweight harness
CUGA (Configurable Generalist Agent) is an open-source agent harness from IBM Research that handles orchestration, planning, tool execution, state management, and guardrails so developers only need to define a tool list and a prompt. The post walks through a working IBM Cloud architecture advisor app built as a single FastAPI file, demonstrating how inline Python tools and shared MCP servers combine with a CugaAgent constructor. Key features include CodeAct execution, reflection-based self-correction, six declarative policy types (Intent Guard, Tool Approval, Tool Guide, Playbook, Output Formatter, CustomPolicy) with semantic matching via sqlite-vec, multi-agent delegation via CugaSupervisor over A2A, and on-the-job skill refinement via ALTK-Evolve. A library of two dozen single-file example apps covers research, productivity, RAG, ops monitoring, and multi-agent patterns. The same agent definition runs air-gapped inside IBM Sovereign Core with per-tool approval, OpenTelemetry tracing, and no external telemetry.