Agentforce 360 Explained: Builder, Agent Script, Voice and Intelligent Context

Agentforce 360 is the name Salesforce now uses for the whole agentic stack: the Agentforce Platform, Data 360 (formerly Data Cloud), the Customer 360 apps and Slack, presented as one layer rather than four products. It was announced at Dreamforce 2025 and reached general availability in early 2026.

If you have been away from this for six months, the version numbers are the least of what moved. The build experience changed, agent topics got renamed, a scripting language arrived, and the model roster grew. Here is what is actually in the box, and — more usefully — which parts you should reach for in which situation.

Last verified against Salesforce Developer documentation on 13 September 2026.

The four capabilities that define the 360 generation

Agentforce Builder

A conversational workspace for building agents. You describe what the agent should do and the builder assembles the configuration, rather than you filling in a form per topic and per action.

This is the on-ramp. It is genuinely fast for a first working agent and genuinely unsuited to anything with a compliance requirement, for reasons the Agent Script section below makes clear.

Agent Script

A compiled, declarative language for agent orchestration, generally available since early 2026 and published as an open, schema-driven language. Its defining idea is hybrid reasoning: deterministic, rules-based execution that the LLM cannot override, sitting in the same engine as generative reasoning that it can.

Scripts compile into an Agent Graph, which is what the Atlas Reasoning Engine actually consumes. Syntax is block-based and indentation-sensitive, closer to YAML or Python than to Apex:

config:
  name: Returns Agent

topic Returns:
  actions:            # deterministic — the script decides these
    - check_order_status
  reasoning:
    actions:          # available to the LLM to call when it judges it should
      - draft_apology_email

The distinction between topic.actions and topic.reasoning.actions is the whole product in miniature. The first list is executed because your logic said so. The second is offered to the model as options. Getting the line between those two right is the core skill.