Most agent frameworks bury you in glue code — tool schemas, JSON parsing, transcript bloat. NVIDIA Labs just open-sourced NOOA (Object-Oriented Agents) with a blunt premise: write your agent as a plain Python object. Methods are the actions the model can call, fields are state, docstrings are the prompt, type annotations are the contract. Software engineering, not prompt engineering.
What makes it different
Six model-facing capabilities live on one surface: typed I/O, code as action, a programmable loop, explicit object state, a model-callable harness API, and the sharp one — pass-by-reference over live objects. Tool outputs stay as Python values instead of getting re-serialized into context. On SWE-bench Verified with GPT-5.5, NOOA scores 82.2% using ~1.1M tokens per task versus rivals’ 2.2M. Half the tokens, higher accuracy, no context compaction. It also takes SOTA on Terminal-Bench 2.0 and ARC-AGI-3.
For developers
It’s a model-agnostic open-source framework (arXiv 2607.20709), shipped through the Open Secure AI Alliance. Drop it in to rebuild your own harness and cut token spend on long agent loops.
You Might Also Like
- Addy Osmani Open Sources Agent Skills 19 Workflows That Make ai Agents Code Like Google Engineers
- Agent Action Protocol aap the Missing Layer Above mcp That Actually Makes Agents Production Ready
- Nvidia Nemotron 3 Super 120b Parameters 12b Active the Math Behind the Fastest Open Source Reasoning Model
- Stripe Coinbase and Ramp Built Internal Coding Agents Langchain Open swe Gives you the Same Architecture for Free
- Arc agi 3 Turns ai Testing Into a Video Game and Every Frontier Model is Losing
Top AI Product 
Leave a comment