Key takeaways

  • NVIDIA announced the expanded Agent Toolkit on October 21, 2026. Agent Knowledge Foundation, Kumo model pages, a Kumo Relational prototype, Auto-Ontology and the NemoClaw optimization blueprint were publicly reachable during review.
  • Kumo has four named routes: Kumo Relational, Kumo Tabular, Kumo Forecast and Kumo Anomaly. NVIDIA links open weights for all four and NIM prototypes for Relational and Tabular; the reviewed Kumo Relational request uses model ID kumo-relational.
  • Agent Knowledge Foundation is a catalog of interoperable models and CUDA-X libraries—not a single mandatory platform. NVIDIA says components work with existing data platforms, ontologies, vector databases and agent frameworks.
  • OpenShell adds a runtime and reference-architecture layer, while NemoClaw adds a trace, evaluation and harness-comparison workflow. Neither page proves policy enforcement or quality gains on a buyer’s own agent.
  • No unified Agent Toolkit price, Kumo production API tariff, service level or region matrix appeared on the reviewed pages. Treat the free prototype and 90-day NVIDIA AI Enterprise evaluation as evaluation access, not a production commercial contract.
01

NVIDIA launched a component stack for enterprise agents

On October 21, 2026, NVIDIA expanded its Agent Toolkit around a new Agent Knowledge Foundation. The announced stack combines Nemotron Retriever, the Kumo structured-data family, Auto-Ontology, CUDA-X libraries, the OpenShell runtime, NemoClaw blueprints and tools including NeMo Relay and NeMo Compass. Public product, model, API and blueprint pages were live when AccessAllGPT reviewed them on October 22.

The important boundary is packaging: this is not one model, endpoint or turn-key agent product. The Agent Knowledge Foundation page says teams can choose components and use them with existing data platforms, ontologies, vector databases and agent frameworks. Buyers therefore need to validate each selected model, library, runtime and governance layer—and the interfaces between them.

02

Kumo makes structured prediction available to agents

NVIDIA names four Kumo routes. Kumo Relational handles classification and regression across connected tables; Kumo Tabular handles single-table classification and regression; Kumo Forecast targets short- and long-horizon forecasting; and Kumo Anomaly targets unusual time-series behavior. The catalog links open weights for all four. Relational and Tabular also link hosted NIM prototypes, while public repositories provide Kumo Structured Data Models and Kumo-TS runtimes.

The Kumo Relational prototype is live behind an API-key flow and uses the request model identifier kumo-relational at /v1/structured-data/nvidia/kumo-relational/predictions. Its contract sends task, schema, relationships, context rows and prediction rows rather than a chat prompt. That makes schema quality, temporal leakage, entity relationships and uncertainty handling first-class evaluation concerns.

03

The knowledge layer spans retrieval, ontology and CUDA-X

Agent Knowledge Foundation joins structured prediction to Nemotron retrieval and Auto-Ontology. NVIDIA describes Auto-Ontology as a way to generate a starting semantic layer that domain experts test, refine and certify. Its public NVIDIA repository is Apache 2.0, but an automatically generated ontology should remain a draft: ambiguous keys, undocumented business definitions and stale relationships can still ground an agent in the wrong answer.

The catalog assigns cuDF to data preparation, cuGraph to connected-data analysis, cuML to conventional prediction, cuOpt to constrained optimization and cuVS to vector indexing and search. NVIDIA publishes improvement figures for several components, including fewer incorrect retrieval answers and faster supported workloads. Those are vendor measurements with workload qualifiers, not a benchmark of the combined stack.

04

OpenShell and NemoClaw cover different control problems

OpenShell is the runtime boundary. NVIDIA’s separate October 21 reference architecture combines its open-source runtime with Vera Rubin inference infrastructure, Vera CPUs for agent harnesses, and BlueField-4 plus Sentry for out-of-band monitoring and policy enforcement. NVIDIA also names Linux, Kubernetes, Windows and major-cloud environments. Claims such as millisecond quarantine, 30-times throughput per megawatt and a 20% capacity increase are NVIDIA-reported and require reproduction in the intended architecture.

NemoClaw for Agent Optimization addresses change control rather than runtime containment. Its blueprint uses NeMo Compass to inspect recorded traces, NeMo Eval Author to create tasks from observed failures, and a held-out A/B comparison before accepting a prompt, tool-description or routing change. The live page includes Hermes and LangChain examples. A saved example demonstrates the workflow; it does not establish that the same change improves another agent.

05

Availability is clearer than production pricing

Several entry points are usable now: downloadable Kumo weights, public runtime repositories, a Kumo Relational prototype, an Auto-Ontology repository and a NemoClaw blueprint with Try Now and GitHub links. The Kumo NIM page also documents self-hosted deployment and advertises a 90-day NVIDIA AI Enterprise evaluation. AccessAllGPT did not authenticate or test whether every account and geography receives identical access.

The reviewed pages do not publish one Agent Toolkit SKU, a production Kumo per-request tariff, throughput quota, support level, region list or data-retention schedule. “Free API endpoint” describes the prototype panel, not a guaranteed free production service. Price GPUs and NIM licensing, data preparation, ontology review, retrieval indexes, evaluation runs, observability, security operations and human adjudication before comparing this stack with a managed agent platform.

06

What agent-platform teams should do next

Start with one decision path, not the whole diagram. For a relational-risk use case, freeze a Kumo artifact or endpoint contract, a time-aware dataset split, schema and relationship definitions, and an agent action ceiling. Compare Kumo against a simple task-specific baseline. Separately test retrieval grounding, ontology review and runtime policy failures; do not turn one successful model score into evidence for the other layers.

Adopt a component only when it beats the baseline under production-shaped data and its license and operating model fit. Constrain predictions to reversible recommendations while calibration and drift are uncertain. Wait when pricing, region, retention or support terms are mandatory and unpublished. Reject the architecture when the team cannot independently enforce tool permissions, reproduce critical outcomes or maintain the semantic layer.

07

Copy-ready Agent Toolkit evaluation record

Complete one record for one bounded agent decision. Label NVIDIA measurements as vendor evidence and your runs as observed evidence.

Entries stay in this browser tab and are not submitted to AccessAllGPT. Blank responses are copied as [Unresolved].

Exact prediction or retrieval task, user, downstream action, reversibility, prohibited actions and required human approval.

Kumo model ID or artifact revision, Nemotron route, ontology version, CUDA-X libraries, OpenShell version, NemoClaw tools and integration commits.

Tables, keys, event time, labels, relationships, missing values, rights, sensitive fields, retention and leakage controls.

Time-aware holdout, simple baseline, slice metrics, calibration, retrieval relevance, ontology review, repeated runs and vendor claims kept separate.

Tool allowlist, network and filesystem policy, credentials, policy-denial tests, logging, incident owner, rollback and containment evidence.

Prototype versus production terms, API or NIM price, GPUs, indexes, data work, ontology review, evaluations, monitoring and human handling cost.

Account and region access, artifact download, endpoint quota, support, SLA, retention terms, license review and offline requirements.

Adopt, constrain, wait or reject; evidence links; rollout ceiling; owner; expiry; and model, schema, ontology, runtime or policy changes that force retesting.

Attach exact requests, dataset and schema hashes, model or container digests, policy files, raw outputs, review decisions and cost telemetry. Never use a partner quote as independent validation.

Primary sources

Browse the publication-wide evidence index →

  1. NVIDIA Agent Toolkit Expands, Grounding Enterprise Agents in Company DataNVIDIA Blog · Reviewed: October 21 announcement date; toolkit components; Agent Knowledge Foundation description; named Kumo models; Auto-Ontology; OpenShell and NemoClaw roles; partner adoption statements; retrieval performance claims · Retrieved · Supports: NVIDIA announced the expanded Agent Toolkit on October 21, 2026, naming Agent Knowledge Foundation, Nemotron Retriever, Kumo, Auto-Ontology, OpenShell, NemoClaw, NeMo Relay and NeMo Compass. Adoption and performance statements are NVIDIA or partner claims.
  2. Building an Agent Knowledge FoundationNVIDIA Developer · Reviewed: Component catalog; compatibility statement; Nemotron Retriever, Kumo and Auto Ontology descriptions; CUDA-X library roles; workflow architecture; published performance claims · Retrieved · Supports: The live product page describes components as individually usable with existing data platforms, ontologies, vector databases and agent frameworks. It lists retrieval, structured prediction, ontology generation, data processing, graph, optimization and vector-search components.
  3. Kumo: Open Models for Structured DataNVIDIA Developer · Reviewed: Kumo family scope; Relational, Tabular, Forecast and Anomaly model descriptions; open-weight links; NIM API links; Structured Data Models and Kumo-TS runtimes; ethical considerations · Retrieved · Supports: NVIDIA exposes four named Kumo routes for relational, tabular, forecasting and anomaly workloads. The page links downloadable weights for all four, NIM API prototypes for Relational and Tabular, and public runtime repositories.
  4. Kumo Relational Model by NVIDIANVIDIA Build · Reviewed: Prototype endpoint; request model identifier; task and schema contract; API-key requirement; self-hosted NIM path; NVIDIA AI Enterprise evaluation statement · Retrieved · Supports: The live page offers a free prototype endpoint using model identifier kumo-relational and documents a self-hosted NIM path. It does not display a production per-request price on the reviewed surface.
  5. NVIDIA Launches OpenShell Reference Architecture for Safe AI AgentsNVIDIA Newsroom · Reviewed: October 21 release date; software and hardware layers; supported operating environments; partner list; availability paragraph; throughput, performance, quarantine and capacity claims · Retrieved · Supports: NVIDIA announced a reference architecture spanning OpenShell software, Vera Rubin, Vera CPUs, BlueField-4 and Sentry, with Linux, Kubernetes, Windows and major-cloud environment statements. All performance and enforcement results are vendor claims.
  6. NVIDIA NemoClaw for Agent OptimizationNVIDIA Build · Reviewed: Blueprint purpose; open status; NeMo Compass and Eval Author workflow; Hermes and LangChain examples; held-out evaluation and A/B comparison steps; Try Now and GitHub access · Retrieved · Supports: The live blueprint documents a trace-to-evaluation workflow for changing prompts, tool descriptions or routing logic, with two runnable examples and saved comparisons. It is an evaluation blueprint, not evidence that an agent improved on a buyer workload.

Limitations

This is a first-party public-source launch briefing, not an independent model, retrieval, ontology, agent, security or infrastructure evaluation. AccessAllGPT did not sign in, generate an NVIDIA API key, call a Kumo endpoint, download or hash weights, clone or audit code, run Auto-Ontology, deploy a NIM, install OpenShell, execute NemoClaw, inspect partner deployments, validate licenses beyond displayed records, reproduce any accuracy, speed, capacity or enforcement claim, or obtain production pricing, quota, region, retention or service-level terms. Product pages, repositories, artifacts and commercial terms can change after retrieval.

Disclosures

AccessAllGPT received no NVIDIA account, API access, credits, model weights, NIM license, hardware, benchmark logs, customer data, private briefing, review, payment or compensation for this article. NVIDIA and its named partners did not sponsor, review or endorse it. AccessAllGPT did not test, score or rank the stack. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with NVIDIA or any named partner. Publication-wide relationships are listed on the disclosures page.

Further AccessAllGPT guidance

  1. NVIDIA Releases Nemotron 3.5 Open Models
  2. NVIDIA Launches RTX Spark Windows AI PCs
  3. AWS Publishes a Nine-Stage Agent Evaluation Framework
  4. IBM Reports Learned Memory Raises Agent Reliability
  5. Design an Agent Benchmark That Predicts Production
  6. Human Approval Boundaries for AI Agents
  7. AccessAllGPT Research methodology
  8. Evidence standards
  9. Publication disclosures

Continue the research

Get evidence-led updates for teams making production AI decisions.