Agent quality is not decided by model name alone. Context rendering, action execution, state handling, and termination logic—the harness—can move benchmark scores and token cost by double-digit percentages.
📑Table of Contents
What was announced
On 2026-07-27, NVIDIA Labs released a research preview called NOOA (NVIDIA Labs Object-Oriented Agents).
An agent is defined as a single Python class, with this mapping:
- methods = capabilities
- fields = state
- docstrings = prompts
- type annotations = contracts
Methods whose bodies are ... are filled by an LLM loop at runtime; ordinary method bodies stay deterministic Python.
Open Secure AI Alliance context
Around the same window, NVIDIA also announced the Open Secure AI Alliance, framing open models, harnesses, and defensive tools as cyber-defense assets and listing NOOA as one technical contribution.
What this article covers
This article is not a press rewrite. It covers:
- NOOA’s design principles
- six harness capabilities and pass-by-reference
- how to read self-reported SWE-bench / CyberGym numbers
- how to separate the alliance story from the code
- a comparison table plus trial steps
Audience and tone
The audience is:
- agent implementers
- teams choosing harnesses
- practitioners feeling token pressure in coding agents
Author first-person experience is not filled in, so the tone stays neutral and operational. For adjacent evaluation and retrieval design, see Continuous MLOps Learning and Agent Evaluation and Designing Search for Agents.
Learning outcomes
What NOOA Is — Agents as Python Objects
NOOA is a research preview that pulls definitions that often scatter across prompt templates, tool schemas, callbacks, and workflow graphs into one Python class.
Class-design mapping
The mapping is:
- Methods = capabilities (actions)
- Fields = state (typed and model-visible)
- Docstrings = prompts
- Type annotations = contracts (typed I/O and validation)
- Methods with
...bodies = filled by an LLM loop at runtime - Ordinary bodies = deterministic Python (unit-testable)
Practical claim and third-party framing
The practical claim is simple: diffs, code review, type checking, and refactors use the same tooling as ordinary software. Primary sources are the official Tech Blog, GitHub: NVIDIA-NeMo/labs-OO-Agents (Apache 2.0), and paper arXiv:2607.20709.
Independently, AI Weekly frames the design as “plain Python objects” and positions the six model-facing ideas against graph/DSL harnesses.
Research-preview limits
Limits are explicit: this is a research preview, and pinning directly to main carries breaking-change risk. Production SLA and enterprise support are not established by the blog alone.
Six Harness Capabilities and Pass-by-Reference
The six ideas NVIDIA lists are explanatory variables for token efficiency as well as accuracy.
The six ideas
- Typed input/output — typed arguments and validated returns instead of free-form text
- Pass by reference — live Python object references instead of full serialization of tool results; the model sees bounded previews
- Code as action — the model acts by writing Python (control flow and method calls)
- Programmable loop engineering — the orchestration loop itself is ordinary Python
- Explicit object state — durable state lives on objects, not only in chat history
- Model-callable harness APIs — the model can inspect/manage context blocks and event history
Memory (typed SQLite)
On memory, NOOA promotes agent-curated typed/tag SQLite with:
- relations such as supports / contradicts / derived-from
- reflection for integration
- NVIDIA-reported +11.8 RHAE on ARC-AGI-3 versus file notes
Why pass-by-reference matters to readers
Pass-by-reference is the bridge for many coding-agent users: it is a design-level answer to “tool output pasted in full melts the context window.”
Assumptions and caveats:
- Assumption: objects stay alive in the execution environment
- Caveat: multi-process and pure HTTP tool boundaries need extra design to reproduce the same shape
How to Read the Benchmarks
Numbers are NVIDIA self-reports plus leaderboard comparisons at post time. For adoption decisions, treat them as conditional upper-bound signals.
Key metrics — SWE-bench and token efficiency
| Metric | NVIDIA-reported summary | How to read it |
|---|---|---|
| SWE-bench Verified | GPT-5.5 82.2% (claims above public SOTA 79.2% at post time); Opus 4.6 79.8%; 253-line general-purpose agent, no bench-only prompts | Model- and harness-dependent; SOTA is time-stamped |
| Token efficiency | 29 calls / ~1.1M tokens/task at 78.2% vs comparison harness 66 calls / 2.2M; alternate 29 / 1.3M at 78.6%; roughly half cost | Re-measure on your workload |
| Context compaction | Claimed unnecessary on SWE-Bench; median peak prompt 22–72k vs 200–400k windows | Assumes pass-by-reference |
Key metrics — CyberGym and ARC-AGI-3
| Metric | NVIDIA-reported summary | How to read it |
|---|---|---|
| CyberGym L1 | GPT-5.5 86.8% with network blocked + trajectory cheat checks | Do not ignore network conditions |
| ARC-AGI-3 | GPT-5.5 mean RHAE 50.2%; GPT-5.6-sol 85.1%; under $20/game ($17.85 / $13.3) | Score–cost Pareto claim |
Benchmark sources
Source: NVIDIA Developer Tech Blog (as of July 2026).
Independent reporting
The Hacker News independently reported CyberGym 86.8% under blocked network. Third-party coverage corroborates that the figure was publicized; it does not guarantee reproduction on your stack.
Pre-adoption checklist
- Do you have a re-measure path on your model API and eval set?
- Can you explain that “above SOTA” is a post-time comparison?
- Are network-blocked security-bench conditions understood?
- Will you measure calls/tokens on one real task before believing half-cost claims?
- Is uncritical republication of bench numbers blocked in internal docs?
Open Secure AI Alliance vs NOOA
The framework story and the industry alliance story connect, but they are not the same ticket.
NVIDIA’s framing
NVIDIA’s corporate blog positions open models, harnesses, and defensive tools as cyber-defense assets.
- Cites a July 2026 Hugging Face incident as motivation for open defensive tooling
- Names NOOA on GitHub as an open research contribution
- Named participants include Microsoft, IBM, Red Hat, Hugging Face, CrowdStrike, Elastic, the Linux Foundation, Nous Research, and many others
Independent corroboration
Independent corroboration includes:
- The Hacker News: ~37 members; NOOA as the first named technical contribution; caveats that governance docs and joint repos were not yet public
- Linux Foundation / Red Hat: inaugural partner/member statements spanning agent harnesses in the defense scope
- CrowdStrike / Elastic: independent join announcements
How to split work internally
Implication: installing NOOA is not “joining the alliance.” Keep these on separate tickets:
- (a) research harness experiments
- (b) open-defense policy watching
Comparison Table and Trial Steps
Do not assume a full replacement. Trial whether an object-oriented surface fits your review culture.
Comparison table — definition and actions
| Dimension | NOOA (research preview) | Typical graph/DSL harness | Decision point |
|---|---|---|---|
| Where definitions live | One Python class | Nodes/edges/YAML/schemas | Strength of code-review culture |
| Action representation | Code-as-action (Python) | Tool JSON / node runners | Typing and testability |
| Tool results | Pass-by-reference claim | Often serialized into context | Token pressure |
Comparison table — state, memory, maturity
| Dimension | NOOA (research preview) | Typical graph/DSL harness | Decision point |
|---|---|---|---|
| State | Explicit object state | Chat-history-centric | Observability on long tasks |
| Memory | Agent-curated SQLite | Summaries / external stores | Audit and backup needs |
| Maturity | Research preview, Apache-2.0 | Product/community variance | SLA and breaking-change tolerance |
| Evidence base | Tech Blog / paper / GitHub | Product docs | Presence of primary sources |
Source: NVIDIA Tech Blog and the labs-OO-Agents README (July 2026).
Trial path (from the GitHub README)
uv inituv add "nooa @ git+https://github.com/NVIDIA-NeMo/labs-OO-Agents.git@main"- Run one SupportAgent-style example with a single
...method - Add a small deterministic method and confirm unit tests still work
- If possible, log calls and tokens against your current harness on one task
Do not (trial stage)
- migrate production under SLA assumptions
- republish bench numbers without conditions
- equate alliance membership with product guarantees
- pin unpinned
mainfor long-lived systems without a freeze plan
FAQ
Q1. Is NOOA a LangGraph replacement?
No. It is a different research-preview design. Treat it as an experiment surface for object-oriented definitions and token pathways, not a mandatory full swap.
Q2. Will we hit 82.2% or 86.8% in our environment?
Those are NVIDIA-reported figures. They depend on model, eval harness, and network conditions. Without a re-measure path, treat them as upper-bound signals.
Q3. Why does pass-by-reference matter?
It aims to stop stuffing full tool outputs into prompts, improving token and cache efficiency (per NVIDIA). Distributed tool boundaries need extra design to keep the same shape.
Q4. Do we need Open Secure AI Alliance membership to use NOOA?
No. The code is an Apache-2.0 public research repository. The alliance is a separate industry initiative.
Q5. When is production use reasonable?
Not while it remains a research preview. Keep it experimental until auditability, reproducibility, breaking-change policy, and support boundaries are acceptable.
Q6. Has the author shipped NOOA in production?
This article stays in neutral practitioner voice. First-person lived experience is not filled in, so no personal success claims are made.
Summary and Next Actions
NOOA’s value is less a new model name than two things:
- a design thesis—agents as Python objects, harnesses as drivers of both accuracy and cost
- reproducible public artifacts (Tech Blog / arXiv / GitHub)
The Open Secure AI Alliance is a policy and defensive-stack context; keep it off the implementation ticket.
This week
- Map the six capabilities onto your current harness vocabulary (one-page memo)
- Install via uv and run one example
- Measure calls and tokens on one existing task to test the pass-by-reference hypothesis
- Put alliance statements on a policy/defense watchlist, separate from implementation work
Do not
- republish benches without conditions
- full-migrate production from a research preview
- treat alliance membership as a product warranty
For evaluation loops, also see Continuous MLOps Learning and Agent Evaluation; for retrieval design, see Designing Search for Agents.
Related articles:
Author
krona23
Over 20 years in the IT industry, serving as Division Head and CTO at multiple companies running large-scale web services in Japan. Experienced across Windows, iOS, Android, and web development. Currently focused on AI-native transformation. At DevGENT, sharing practical guides on AI code editors, automation tools, and LLMs in three languages.
🔥 Most Popular
- Claude Pricing: I Tested All 5 Plans — Here's My Verdict (2026)
- Claude Desktop Won't Install? Windows & Mac Fixes That Worked (2026)
- Cursor Pricing 2026: Plans & Real Costs After 3 Years of Pro
- How to Spot and Defend Against Two-Stage Phishing Emails in 2026
- Docker Sandboxes (sbx) Guide: Run Claude Code Safely in a microVM
Leave a Reply