Salesforce Gets Serious About AI Agents: Headless 360 Opens Platform via API

Contents

Salesforce provides core functions of its platform programmatically with “Headless 360.” Data, workflows, and business logic can be controlled directly via more than 60 MCP tools, over 30 predefined coding skills, as well as APIs and a CLI. In parallel, the provider is expanding its development environment with “Agentforce Vibes 2.0” to include AI-powered functions.

With Headless 360, Salesforce is shifting access from the graphical interface to interfaces. Applications are no longer used exclusively through UI interactions but are addressed via APIs, MCP tools, and automated agents.

Technically, the approach is based on an expanded API-first architecture. Functions, data, and permissions are available independently of the interface and can be used via various clients. “When you start with APIs and combine that with Slack and other clients, new agentic experiences emerge,” says Gary Lerhaupt, VP Product Architecture at Salesforce. The company sees this as the foundation for an “Agentic Enterprise,” in which interactions increasingly occur via various clients rather than traditional applications.

Development and Execution of Agents

A central component is the integration of external development environments. Tools like Claude Code, Cursor, or Codex access the platform via MCP and cover phases such as data modeling, implementation, testing, and deployment with the help of pre-configured coding skills. For developers who want full control over the visual layer, the platform also supports native React development.

Additionally, Salesforce provides a browser-based development environment with “Agentforce Vibes 2.0” that integrates AI-powered coding functions. Deployments can be triggered via natural language through the DevOps Center MCP. Salesforce estimates development cycle reductions of up to 40 percent; independent benchmarks are not yet available.

With the “Agentforce Experience Layer,” Salesforce separates business logic and presentation. Interactions such as approvals or workflows only need to be defined once and are then available in various clients, such as Slack, mobile applications, or external AI interfaces.

Tools for Operation and Control

For agent control, Salesforce is introducing the scripting language “Agent Script,” which combines deterministic logic with natural language instructions. It is available under an open-source license and is used for orchestrating multiple agents. Background: According to Gary Lerhaupt, pure LLM approaches achieve only around 95 percent reliability, but productive enterprise systems require at least 99.5 percent.

The “Testing Center” (available from May 2026) is intended to detect logic gaps before deployment. “Custom Scoring Evals” evaluate decisions, “Observability” and “Session Tracing” analyze interactions. An A/B testing API allows parallel testing of different agent versions.

Practical Examples from Companies

As a practical example, Salesforce cites the booking platform Engine. According to the company, they have put a customer service agent into production within twelve days, which handles about half of the inquiries automatically. The prerequisite was the preparation of historical data.

Organizational adjustments also accompany the implementation. The company asymbl supports companies in integrating digital workforces into existing processes and operates numerous AI agents, according to its statements. Ongoing operation requires continuous monitoring and readjustment - “There is no set and forget. It requires intention, method, and ultimately coaching,” says CEO Brandon Metcalf.

The innovations show a strategic shift: Salesforce is expanding its platform into an infrastructure for agent-based systems. It remains open how stable these will operate in productive use and how the increasing technical complexity will affect development and operation. Furthermore, the benefit strongly depends on the quality of the underlying data - a point that the Engine example also underscores, where the preparation of historical data was a prerequisite for productive operation.

(afl)