1. Salesforce Agentforce 360: AI Agents with Context

• Trusted Context as a Foundation
• Hybrid Reasoning for Controllable Processes
• Technical Advances with Limitations

Salesforce is expanding its Agentforce 360 platform with new features aimed at addressing the much-discussed context gap of today’s AI agents. In addition to tools like Agent Script, an enhanced Agent Builder, and voice integration, the focus is on closer integration with the acquired Informatica technology. This forms the basis for the company-wide context that agents will access in the future. In parallel, Salesforce is opening up its platform, allowing partners to develop and market their own products directly based on the Agentforce architecture.

Trusted Context as a Foundation

At the core of the announcements is a consolidated data architecture, which Salesforce describes as a prerequisite for reliable agent approaches. The Data 360 platform will be more closely integrated with the acquired Informatica technology and supplemented with real-time signals from MuleSoft. The goal is to create a metadata framework that brings together master data, catalogs, data lineage, and operational events. Agents will no longer just generate text but will be able to access clearly defined business objects and correctly interpret their state.

Peter Wüst, Senior Vice President Solution Engineering, describes this combination as a "context engine" that no longer allows models to work with fragmented data views. While AI models are powerful, they are "enterprise-dumb" because they lack specific knowledge about a business, its history, and its rules. Without this shared understanding, agents are ultimately forced to guess – a core problem that Salesforce aims to address with the new data foundation.

The data foundation also forms the basis for deterministic agent logic, which Salesforce classifies as central to reliability and compliance.

Hybrid Reasoning for Controllable Processes

One of the biggest problems with large language models in enterprise use remains their lack of predictability. Salesforce is trying to mitigate this through Hybrid Reasoning: while the LLM continues to be responsible for interpretation, process execution is outsourced to fixed logic modules. The new scripting language, Agent Script, is used to model these steps and link them with existing Salesforce flows. This is intended to enable agents to execute tasks in a structured and reproducible manner, rather than relying solely on probabilistic model decisions.

In parallel, Salesforce is supplementing its platform with tools to make agents easier to use. The revised Agent Builder generates agent logic from naturally formulated instructions and is intended to simplify the modeling of complex processes. With Agentforce Voice, whose rollout has been repeatedly delayed and which will also support German at the beginning of 2026, a deeply integrated voice interface is added. It works bidirectionally, multilingual, and can be used in CRM, Service, and Commerce applications. Wüst describes how everyday tasks can be simplified with it: Field service employees can dictate visit reports immediately after an appointment, whereupon the system structures the content, updates data records, and initiates follow-up processes if necessary. Salesforce sees this combination of voice input and business logic as a building block for making agents more widely usable in everyday work.

For the first time, Salesforce is also allowing partners to integrate components such as Agentforce 360, Data 360, and Trusted Services directly into their own products and sell them commercially. This is accompanied by more flexible usage models and an expanded marketplace designed to automate provisioning and billing. The opening aims to make agent functions more widely available not only in Salesforce applications but throughout the entire ecosystem.

Technical Advances with Limitations

Salesforce is pursuing a clear line with this: agents are to evolve from experimental tools to regulable building blocks in enterprise operations. The combination of data context, deterministic processes, and voice connectivity addresses known weaknesses of current AI solutions. However, it remains open how much effort companies will need to achieve the necessary data quality – and how much the dependency on the tightly integrated platform will increase. It will be crucial whether Salesforce can maintain the balance between integration depth and openness.

(mack)

Don’t miss any news – follow us on Facebook, LinkedIn or Mastodon.

This article was originally published in German. It was translated with technical assistance and editorially reviewed before publication.