If an experimental agent framework can cut total token consumption in half and reduce foundation model invocations by more than 50% while scoring higher on software engineering benchmarks, is performance primarily a property of the model, or does it belong to the execution harness?
That question emerged from Nvidia’s release of Noah, an experimental agent framework that models autonomous workflows as strongly typed Python classes. Instead of relying on open-ended conversational prompts, Noah treats class docstrings as system prompts, class instance fields as persistent state variables, and generated Python code as the workflow orchestration engine.
On SWE-bench Verified, Nvidia reported that Noah reached an 82.2% resolution score using approximately 28 model calls and 1.1 million tokens per task. By contrast, a leading competing harness consumed 66 model calls and 2.2 million tokens per task while scoring four percentage points lower. The dramatic reduction in API calls and token volume proves that agent efficiency is an architectural property of state representation and execution design.
Nvidia Noah: Halving Tokens and Calls Through Typed Class Agents
Traditional agent frameworks suffer from conversational drift. When an agent interacts through free-form natural language, each intermediate thought and tool output appends directly to the context window. The prompt expands rapidly, degrading model attention and forcing the system to re-read thousands of irrelevant tokens on every turn.
Noah replaces unstructured conversational history with object-oriented state management:
- Typed state fields: The agent updates specific attributes on a Python class, discarding intermediate conversational noise.
- Deterministic control flow: The agent writes executable Python control flow rather than generating sequential JSON tool calls, handling loops and conditions locally.
- Focused execution: By maintaining state inside strongly typed variables, the model executes complex tasks in fewer, more targeted turns.
While cutting token consumption by 50% is a major operational achievement, Nvidia’s benchmark figures remain vendor-reported and require independent verification across external workloads. Nevertheless, the architecture demonstrates that structuring how a model receives state is the most effective lever for optimizing agent economics.
Type Validation Is Not Authorization: The Containment Gap
While strongly typed classes provide elegant programming abstractions, developers must not confuse output schema validation with operational authorization. Noah uses Python type checkers to verify that a model’s generated output conforms to expected structural shapes.
However, confirming that an output matches an expected data structure provides zero assurance that the requested action is safe or authorized:
- Schema validity versus business logic: A model generating an administrative refund may produce a syntactically valid JSON payload, but that schema check cannot determine whether the user is authorized to receive the funds.
- Arbitrary code execution: Because Noah executes generated Python scripts directly, the framework explicitly warns that its internal validators do not constitute a security boundary.
- Isolation requirements: Operating Noah safely requires enclosing the runtime within ephemeral virtual machines or hardened containers with restricted network access.
Confusing syntactic validation with authorization security invites severe production vulnerabilities. Runtimes must enforce out-of-band authorization policies and immutable audit logging before executing state-changing operations.
VRAM Budgeting for Ornith 1.5: When 170k Contexts Degrade at 80k
The distinction between advertised specifications and usable performance is equally apparent in local model execution. Ornith 1.5 35B packages 36 billion total parameters while activating approximately 3 billion parameters per token. In documented desktop benchmarks on an Nvidia RTX 4070 Ti, the model generated 50 to 56 tokens per second while sustaining prompt processing speeds between 650 and 700 tokens per second.
However, allocating maximum context windows creates severe hardware and cognitive bottlenecks. The 21.71 GB Q4KM quantized build barely fits inside the 24 GB VRAM boundary of an RTX 4090, leaving minimal video memory for active key-value caches.
Furthermore, allocating a 170,000-token context window does not mean the model can reason across it effectively. Empirical evaluations of Ornith 1.0 revealed that retrieval accuracy and logical coherence deteriorated rapidly once filled context exceeded 70,000 to 80,000 tokens. Allocating memory for context that degrades reasoning is an expensive engineering mistake. Context size must be benchmarked against factual retention curves rather than theoretical limits.
Stripe Acquires OpenRouter: The Consolidation of 25 Trillion Weekly Tokens
As autonomous systems multiply API calls, model gateway infrastructure is undergoing rapid commercial consolidation. Financial infrastructure giant Stripe acquired OpenRouter, merging global payments rails with a model routing gateway that connects developers to more than 400 foundation models across over 80 independent hosting providers.
In May 2026, OpenRouter was routing an astonishing 25 trillion tokens per week. For engineering teams, the platform provides unified billing, standardized API abstractions, and automated failover across competing model providers.
However, consolidation introduces significant platform risk. By owning the gateway, Stripe controls the routing layer, the usage metering, and the underlying payment settlement. An engineering abstraction that simplifies multi-model access simultaneously creates a centralized point of commercial control. Building resilient AI infrastructure requires maintaining direct provider adapters alongside third-party routing aggregators.