NVIDIA dropped NOOA - an object-oriented agent framework that scored 82.2% on SWE-Bench Verified while using 50% fewer tokens than competing models.
The architecture treats code entities as objects with state and methods, which apparently makes the agent's reasoning more structured and token-efficient. SWE-Bench Verified is the harder subset of the benchmark (only 500 verified issues vs 2,294 in full set), so 82.2% is legitimately strong.
Key technical win: half the token cost means faster inference and lower API bills for the same quality. Framework is open source now, so you can actually dig into how they structured the agent loops and object representations.
If you're building agentic code systems, worth checking their approach to state management and task decomposition.
The architecture treats code entities as objects with state and methods, which apparently makes the agent's reasoning more structured and token-efficient. SWE-Bench Verified is the harder subset of the benchmark (only 500 verified issues vs 2,294 in full set), so 82.2% is legitimately strong.
Key technical win: half the token cost means faster inference and lower API bills for the same quality. Framework is open source now, so you can actually dig into how they structured the agent loops and object representations.
If you're building agentic code systems, worth checking their approach to state management and task decomposition.