IBM's CUGA cuts agent building down to tools and a prompt

IBM has released CUGA, an open-source harness for building AI agents, along with a repository of 24 example apps that each fit in a single file. The framing is deliberate: CUGA is a harness, not a framework. It supplies the planning loop, tool calls, state tracking, and error handling that you would otherwise rebuild for every agent, and asks you to write only the tools and the prompt. The IBM Cloud advisor example is four arguments to a CugaAgent constructor.

Under that small surface is real machinery. CUGA plans with reflection, so it re-plans when a run goes off course instead of pushing ahead, mixes generated code with tool calls, and keeps variables straight across runs of twenty or more steps. Tools return a uniform envelope, a success or an error object, so a failing tool degrades gracefully rather than derailing the agent. There are three cost and latency modes you can switch without touching code, and the same agent runs locally, in Docker, or in a cloud sandbox.

The benchmark claim is the interesting part. IBM Research writes that CUGA topped AppWorld and led WebArena, and that it does so partly by moving orchestration work out of the model, which let a smaller open-weight model like gpt-oss-120b compete with frontier systems. Governance sits in the runtime too, with intent guards, human approval for risky tool calls, and pinned playbooks for known-good procedures.

Why it matters

If you keep rewriting the same agent plumbing, CUGA is worth a look: the examples show a working agent can be one file, and the benchmark results suggest good orchestration lets a cheaper open model do work you might assume needs a frontier one. That is a direct lever on your agent's cost.