NVIDIA's NOOA represents AI agents as plain Python objects, ditching prompt/config sprawl
burkov · x · 2026-09-03
AI agent behavior is usually spread across prompts, tool definitions, callbacks, config files, and workflow graphs. NVIDIA researchers propose NOOA (NVIDIA Object-Oriented Agents), exploring whether much of that machinery can simply become ordinary Python.
The core design represents an agent as a Python object: methods are actions, fields hold state, docstrings contain instructions, and type annotations constrain inputs/outputs; some methods run normal deterministic code while others are filled in at runtime by an LLM. The model can also write Python to inspect and manipulate real program objects instead of repeatedly converting everything to text. The paper offers a practical lens on a question model research often ignores: how much agent reliability comes from the model versus the engineering scaffolding.
More from coding & agent
- HarnessEvolve paper: dual-gate loop fixes three failure modes of self-evolving agents — dair_ai · 2026-09-03
- Mastra launches sandbox computer use so agents can browse, click, type and screenshot — Scobleizer · 2026-09-03
- How to Keep Long-Running Agents on Track: Redis State Beats Context Stuffing — Deepfeet-09 · 2026-09-03
- Looking for a <500M SLM to Summarize Code Snippets for Local Agents — Mrinohk · 2026-09-03
- LangSmith Roadshow heads to Dallas in two weeks with agent dev workshops — LangChain · 2026-09-03
- Agentic API adds a stateful layer in front of vLLM for open-model agent runtimes — techNmak · 2026-09-03