Enterprise agents grow up: Oracle's Claw, Google's policy CLI, Dust's multi-model layer

Oracle packages the agent

Oracle made the biggest enterprise-agent packaging move of the day. Its new AI Agent Studio for Industry Applications is a no-code and low-code environment for creating, testing, orchestrating, deploying and managing AI agents for industry-specific processes, with MCP server support, monitoring and observability for agent execution metrics, and guardrails that keep agents within configured permissions, policies and organizational roles. It targets nine industries - engineering, construction, retail, finance, restaurants, government, utilities, communications and hospitality - and is planned to be available within the next 12 months, starting in the US before expanding to other OCI Enterprise AI regions. m:…0042

Alongside it, Oracle introduced Claw-powered agentic applications for Fusion HCM. Claw is a governed agentic execution runtime - the company says it lets Fusion Agentic Applications reason, compute, adapt and execute specialist-grade work, from deep research and modeling to continuous re-planning, within established governance controls. The HR apps include Workforce Orchestration, which develops plans balancing demand shifts, cost and operational continuity, plus Candidate Sourcing, Recruiting Command Center, Payroll Command Center and Workforce Compensation modules. m:…0566

Google governs the build

Google took the developer side of the same problem. Its Agents CLI Extension System lets developers use LangChain, LangGraph or custom enterprise stacks while keeping a unified deployment and evaluation workflow - including a first-party LangChain template that scaffolds, runs and deploys agents without requiring Google's Agent Development Kit. The governance hook, the company says: platform teams can author custom extensions to enforce policies such as automated software bill of materials (SBOM) scans during deployment. m:…6322

The control layer

The independent layer came from Dust, spotlighted by Snowflake's startup program. Dust is a collaborative AI platform that lets teams switch among supported frontier and open-source models, captures agentic traces - instructions, context, model and tool choices, outputs and human feedback - and lets them govern the context, agents and traces their work generates. Co-founder Stanislas Polu's argument, via the Snowflake piece: adopting AI one assistant or model at a time creates dependency on a single provider, loss of control over agent-generated data, and no visibility into costs and behavior. His operating principle: "Agents propose; humans dispose." Dust says it deliberately avoids training models to focus on the product layer where customers have concrete problems, and its agents can work with Snowflake structured business data combined with context from CRM, Slack, Notion or call notes. m:…4535

NVIDIA's expanded open-source Agent Toolkit fills the data-grounding half. The Agent Knowledge Foundation turns company data into intelligence agents can use without moving data, rebuilding workflows or changing governance controls; the new Auto-Ontology library translates raw company data into familiar business concepts - customers, products, orders - and maps their relationships. The company cites early enterprise results: SAP's fine-tuned reranker ran nearly 3x faster with Nemotron Retriever, and ServiceNow improved reranking quality while cutting latency by nearly a quarter. m:…3606

The pattern

Our read: the agent conversation has moved from what models can do to how enterprises run them. Packaging (Oracle), policy enforcement (Google), model independence (Dust) and data grounding (NVIDIA) are four answers to the same buyer question - who is in control when agents act. The vendors competing on governance controls, not raw capability, are the ones winning enterprise budgets. m:…0566 m:…6322 m:…4535 m:…3606