Nvidia’s NOOA makes an agent one Python class

Nvidia's NOOA (Object-Oriented Agents) introduces a novel approach to agent development, proposing that an agent is essentially a single Python class. This class encompasses an agent's capabilities (methods), its state (fields), and prompts (docstrings). The Python class body is replaced with an LLM-driven loop when it contains an ellipsis (…).

This design aims to address the fragmentation issue that has plagued agent development, where various components such as prompts, tool definitions, and callbacks are scattered across different abstractions and languages.

Nvidia's chief AI architect, Adnan Masood, believes that this approach could simplify agent testing and tracing, making it more akin to traditional software development. However, other experts remain skeptical. IBM's solution architect, Karthik Karunanithi, questions whether this centralized approach will introduce new review problems, particularly with deterministic and probabilistic behaviors intermixed within the same method signature.

Siddhartha Saxena, co-founder of Thine and Merlin AI, praises the use of typed input/output in NOOA for providing structure to agent calls. Yet, he cautions that scaling to millions of tool calls may present testing challenges, viewing them more as observability concerns rather than solvable issues by the framework itself. Both centralization and readability may introduce new risks, with potential security implications due to the model's ability to act upon and execute Python code.

Nvidia's benchmark results suggest that harness design can significantly impact performance, with NOOA achieving parity or better, at roughly half the cost, compared to traditional harnesses. However, Karunanithi warns against overgeneralizing these results, emphasizing that the tested benchmarks differ from production systems where security and regulatory compliance are paramount.