NVIDIA-labs OO Agents: Native Python Object-Oriented Agents
NOOA treats an agent as a live Python object, with methods as actions and fields as state.
The paper says LLM-completed methods can sit beside ordinary deterministic Python methods, using docstrings as prompts and type annotations as contracts. NVIDIA’s authors argue this gives developers and agents the same interface for testing, tracing, refactoring, and improving behavior. They also claim the framework combines typed I/O, pass-by-reference over live objects, code as action, explicit state, loop engineering, and model-callable harness APIs on one surface. The authors report tests on targeted capabilities and benchmarks including SWE-bench Verified, Terminal-Bench 2.0, and ARC-AGI-3. HF Daily Papers' note
The paper says LLM-completed methods can sit beside ordinary deterministic Python methods, using docstrings as prompts and type annotations as contracts. NVIDIA’s authors argue this gives developers and agents the same interface for testing, tracing, refactoring, and improving behavior. They also claim the framework combines typed I/O, pass-by-reference over live objects, code as action, explicit state, loop engineering, and model-callable harness APIs on one surface. The authors report tests on targeted capabilities and benchmarks including SWE-bench Verified, Terminal-Bench 2.0, and ARC-AGI-3. HF Daily Papers' note
score 6