AI rules instead of black box: Deterministic control of AI agents

Contents

Salesforce is expanding its Agent Fabric platform with features for integrating, managing, and controlling AI agents. This allows multiple agents across different systems and platform providers to be centrally controlled.

The central innovation for orchestrating AI agents is the integration of Agentforce Script into Agent Broker. With this, Salesforce is transferring so-called guided determinism to cross-platform multi-agent control. Developers define their handoffs as fixed code, with language models handling processing in between.

All multi-agent workflows can be controlled via a graphical interface using drag-and-drop, and human approval points can be viewed. There is also integration with MuleSoft Vibes for controlling workflows in natural language. Independent data on the broker's performance is not yet available. Salesforce uses so-called “Agentic Work Units” for billing. They record usage and interactions, but not business value.

Governance and Integration for Agents

With AI Gateway, Salesforce is introducing a central instance for controlling language models. Routing between models, token budgets, and security and compliance requirements can be defined in one tool. The platform remains model-agnostic. In addition to predefined models, companies can connect LLMs of their own choice. Besides models from OpenAI, Anthropic, or Google, their own language models can also be used.

Furthermore, the platform receives MCP Bridge, an interface for integrating existing systems. It converts existing REST APIs into MCP endpoints without code changes. Agent Fabric also handles authentication and rate limiting. Additionally, Salesforce is introducing Trusted Agent Identity, a permission model for AI agents. Actions are performed within the context of specific user rights. For critical operations, such as financial transactions or legal approvals, confirmation via smartphone is also required. Agent Fabric also logs all actions.

External Agents in the Registry

Agent Scanners capture external AI agents and add them to a central registration database. Via OAuth, they now support agents from Amazon Bedrock, Microsoft Foundry, and GoDaddy. A so-called “Controlled Registration” regulates which agents the tool adds to the registry. MCP servers also provide data quality and governance functions directly from the registry. All registered agents, along with their policies and model usage, can be viewed via dashboards.

Salesforce cites an integration with the instant messaging service Slack as a practical example. Here, users enter a prompt, and Agent Fabric then selects suitable agents or external systems via MCP interfaces. The US company also lists other possible use cases. For example, users should be able to prepare workshops with AI support, create software requirements, or receive assistance with development.

Stronger Focus on Multi-Vendor Environments

With the extensions in Agent Fabric, Salesforce is transferring existing concepts from API management to agent systems, including Gateway, Registry, and Policy Engine. At the same time, the company wants to focus more on multi-vendor environments, explains John Kucera, SVP Product Management at Salesforce. Customers would not rely on a single agent platform.

Nevertheless, Agent Fabric remains closely tied to the Salesforce infrastructure through the MuleSoft infrastructure and the Agentforce script format. The maturity of the MCP and A2A protocols, both less than two years old and implemented inconsistently between providers, remains unclear.

AI Gateway, MCP Bridge, and Trusted Agent Identity are generally available immediately. Deterministic orchestration in Agent Broker will start as a beta in April 2026, with a visual authoring canvas and support for Salesforce models, it should be available to all customers from June 2026.

Agent Scanners for Amazon Bedrock, Azure AI Foundry, and GoDaddy are available now; MCP server support will follow in May, and broader OAuth integration in June 2026. Agent Fabric is also now available in the Canada and Japan server regions and allows operation on MuleSoft Runtime Fabric in private cloud and on-premises environments.

(vbr)