As organizations move from experimental AI to production-grade agents, the requirements shift from “making it work” to “making it secure, observable, and scalable.” Unlike traditional, deterministic software, AI agents operate in a probabilistic manner and with various degrees of autonomy, requiring a fundamental shift in platform engineering approach.
To help organizations navigate this shift, at Explore 2026, Broadcom is announcing enhanced capabilities in VMware Tanzu Platform centered around five critical pillars for AI agents:
- Secure-by-Default Architecture: Enforcing strict agent containment using a deny-by-default runtime and separate credential storage to prevent sandbox breakouts, reducing prompt injection attacks. We are also announcing identity for agents running on the platform so that you can attribute each agent to their individual actions.
- Governed & Advanced Patterns for Agents: Curated marketplace to centralize compliant services, tools, and MCP servers so teams can safely discover, publish, and bind resources via single-command without risky workarounds.
- Streamlined Agent Delivery: Accelerates deployment with enhancements to the Agent Buildpack with an out-of-the-box harness, persistent memory, and customizable human-in-the-loop controls for regulated autonomy. We are also introducing progressive skills disclosure for improved context window efficiency.
- Auditable & Controlled Agents: An integrated AI Gateway tracks per agent audit metrics in Tanzu Hub. View agent tool calls, prompts, resources read as well debug failures, for audit trail and ongoing agent optimization. Visibility of per agent token usage and quotas helps avoid costly runaway agent argument loops, while tool filtering constrains agents to more defined paths.
- Sovereign & AI-Ready Data: New AI-ready data foundations in Tanzu Platform enable multimodal data ingestion, parsing and semantic layering turning raw enterprise data into isolated, AI-ready data products that agents can safely query without exposing sensitive assets.
1. Safer, more secure agents with secure-by-default architecture
This year, we introduced Tanzu Platform agent foundations specifically for agentic workloads. Tanzu Platform agent foundations provide a deny-by-default architecture, enabling complete agent containment. Each agent operates within its own secure sandbox and is only allowed access to models, tools, and data through unique credentials. Access to models and tools is explicitly granted via brokered service bindings with credential management, to prevent agents from accessing unauthorized internal systems. In this way, organizations can maintain better control of their agents and reduce the risk of agent breakouts.
While autonomy is a core tenet of agent design, restricting resources too severely can starve an agent, preventing it from fulfilling its mission. To successfully execute complex tasks, agents require access to various tools, data products, and sub-agents. Traditionally, this necessitates providing credentials directly to the agent, which introduces significant security vulnerabilities if mishandled by an autonomous entity. To mitigate these risks, Tanzu Platform will implement a dedicated, separate credential store positioned outside the agent’s secure sandbox. This architectural isolation prevents agents from directly viewing or leaking its credentials, effectively thwarting prompt injection attacks and preventing malicious actors from exploiting the agent to harvest enterprise credentials.
In addition, Tanzu Platform will also include modern agent identity capabilities to establish clear lineage by tracing agent actions back to their origin and attributing them directly to individual employees. For agents hosted on Tanzu Platform, organizations will be able to track every agent, significantly streamlining root-cause analysis and improving Mean Time to Repair (MTTR). Furthermore, agents can be configured to inherit only the specific permissions of the initiating user, preventing them from taking actions that exceed human authorization and mitigating the risk of rogue behavior.
2. Advanced and compliant agent integrations with a centralized marketplace
Fragmented, off-platform workarounds are the enemy of enterprise scale. Tanzu Platform provides a centralized marketplace where platform engineering teams can curate and better secure services. The marketplace acts as a governed library that enables teams to build more consistent agents. Developers, or agents, can seamlessly discover approved services, including AI models running on the platform and securely bind to them directly with a single command.
The Tanzu Platform marketplace fosters cross-team collaboration by allowing developers to publish custom tools and services, making innovations reusable across the organization. Following the introduction of MCP server publishing in Tanzu Platform 10.3—which enhanced agent tool visibility and lifecycle management—Tanzu Platform will support publishing of agents (custom, pre-built and third-party) to the marketplace. These vetted agents can be consumed by other teams or integrated into advanced, multi-agent orchestration patterns. Furthermore, every published agent is fully observable, with telemetry shared directly in Tanzu Hub for continuous optimization.
To further enhance enterprise governance, we are announcing that Tanzu Platform marketplace can streamline the distribution of curated data products. Curated data products enhance strict security perimeters by distancing the agent from the raw backend systems. These isolated, sovereign data assets also give agents organizational, semantic knowledge, enabling them to make more accurate decisions. By supplying curated data products that sharpen agent focus, organizations can fundamentally increase the accuracy and reliability of agents and improve token usage, enabling agents to ingest only the critical data necessary for their specific mission. Consequently, enterprises can reduce hallucinations through improved context while controlling token usage.
Figure 1: Tanzu Platform marketplace now includes data products for agents – separate, AI-ready pre-processed data slides made available to agents
3. Streamlined agent delivery with the Tanzu Agent Buildpack
Tanzu Platform continues to be harness-agnostic which provides organizations with more flexibility for building and running agents. However, we do offer a turn-key, low-code agent harness, delivered via the Agent Buildpack that makes it easier to deploy an agent, drastically cutting agent development cycles so you can ship production-ready agents faster.
In Tanzu Platform 10.4, we announced Technical Preview of the Agent Buildpack which included a pre-built agent loop and the ability to bind to AI models, agent skills (limited) and connect to MCP servers for rapid, out-of-the-box agent deployment. At Explore 2026, we are introducing enhancements to the Agent Buildpack which include new features that help to create more powerful agents that run scripts and interact with the agent’s filesystem, enabling more autonomous problem-solving agents that can execute complex multi-step workflows.
We have also expanded capabilities for attaching short and long term memory services for agents, backed by persistent memory. By automating connection of memory service to any agent type, the agent memory service helps to improve agent response accuracy while decreasing agent response times (less looping). The provided memory service includes both a REST API and MCP server to easily connect to any standard agent. By storing the agent’s unique memories, you can improve agent accuracy and also enable agents to share memories with each other in agent-to-agent (A2A) interactions. When agents connect to other agents via the Tanzu Platform marketplace they can take their unique memories with them.
In addition to the agent memory service, we will also be adding enterprise controls for agent autonomy in the Agent Buildpack. Organizations can set levels of human-in-the loop controls for agents that configure each agent’s level of autonomy. Custom permission levels can be set to “always deny,” “always allow” or “request before taking action” for agentic workflows. We have also added progressive disclosure capabilities to help agents progressively load skills rather than loading the entirety of each skill. This increases the agent’s accuracy and efficiency by reducing the amount of information that floods the context window.
4. Auditable and controlled agents via AI gateway
Agents are designed to operate with autonomy but at machine speed. This makes traditional logging insufficient for the speed and breadth of an agent’s reach. Yet, observing any application behavior remains a required enterprise audit practice and an important risk mitigation strategy. Tanzu Platform addresses agent observability by intercepting all agent traffic through a robust AI gateway. Our AI Gateway provides platform engineering teams with the ability to see which tools an agent uses and tune those tools accordingly. Tanzu Platform AI gateway tracks per agent audit metrics in Tanzu Hub. Each agent’s actions are tracked and logged to provide a complete audit trail. You can view tool calls, prompts, resources read as well as debug failures in the audit log so you can perform ongoing agent tuning. Platform engineering teams can also observe model responses and export journals for model interaction analysis through the AI Gateway, enabling models to perform optimally. By setting token rate limits for model usage and exporting journals for interaction analysis, teams can avoid runaway agent loops and ensure models perform optimally.
At Explore 2026, we are introducing tool filtering capabilities in the AI gateway so that Platform Engineering teams can limit the number of tools that are exposed to agents. Platform teams are able to define (regex) rules to enable or disable tools for an MCP server. Any tool that matches a disabled rule is not exposed to clients connecting through that gateway endpoint thereby limiting the agents from non-approved behavior.
5. Sovereign data for agents with Tanzu Platform AI-ready foundations
Agents are only as good as the data they access. Tanzu Platform enables developers to build more secure, sovereign data pipelines for agents, by providing pipelines to parse and structure data directly on the platform. By offloading complex, multi-modal data processing from the agent, organizations can increase reliability and performance. At Explore 2026, we announced AI-ready data foundations to help organizations to transform chaotic data into governed data products. Read the blog to learn more.
Scaling Trusted Agents in the Enterprise
Building the future of the agentic enterprise requires more than just model hosting; it demands a robust, integrated runtime and institutional data that prioritizes security and governance from the ground up. By combining secure-by-default architecture, centralized marketplace services, granular observability with multi-modal data transformation, Tanzu Platform delivers the operational substrate required to transition from experimental AI to scalable, production-ready agents. Ultimately, this comprehensive approach empowers organizations to harness the full potential of autonomous agents while maintaining the control necessary for modern enterprise environments.
Read more about AI agent use cases enabled by Tanzu Platform or or contact the Tanzu sales team
