TL;DR — Key Takeaways
- Salesforce’s Enterprise AI Harness is designed to provide a common management and governance layer for AI agents operating across Salesforce and third-party platforms.
- The architecture spans six areas: context, action, agency, security, governance and models.
- Salesforce says the software surrounding an AI model can have a major impact on performance, with its research showing task success rising from 29.2% to 78.0% without changing model weights.
Salesforce has unveiled Enterprise AI Harness, a collection of technologies designed to manage and govern AI agents across enterprise systems, including agents and models from third-party providers.
A key challenge for enterprise companies is that AI agents are being deployed across multiple platforms, yet they need access to corporate data and applications while operating within security and governance policies. Salesforce’s strategy is to provide a common layer that controls how these agents obtain information and perform tasks.
Many enterprise companies operate multiple agent platforms, and this multi-vendor environment is creating a new competitive battleground. ServiceNow offers AI Control Tower, while Microsoft, AWS and other vendors have developed technology for managing AI agents. Salesforce previously entered this market with MuleSoft Agent Fabric, which was previewed at Dreamforce last year and released in January.
Enterprise AI Harness expands that effort by drawing on technology from Salesforce products and acquisitions including MuleSoft, Informatica, Data 360, Tableau and Agentforce. Customers will be able to use Salesforce components alongside third-party agents, models and existing enterprise systems.
Six Areas For Enterprise AI
The Enterprise AI Harness architecture is divided into six areas: context, action agency, security, governance and models.
This approach reflects a key shift in enterprise AI: A capable model alone is not enough to safely automate business processes. An agent may need information from CRM and ERP platforms, contracts and previous customer interactions before taking an action. It also needs rules defining which data it can see and which actions it can execute.
Salesforce research purports to show how much the software surrounding an AI model can affect results. Company researchers tested a technical agent harness across seven enterprise benchmarks. Changes to the harness around a smaller Qwen model raised average task success from 29.2% to 78.0% without changing the model weights. A subsequent fine-tuning approach reduced success to 63.1%, while a more targeted training technique lifted it to 79.7%.
The results also suggest that swapping models inside an enterprise AI system may not always be seamless because model performance can depend on how the surrounding agent architecture is configured.
Salesforce is also introducing an AI Control Plane that allows IT teams to register agents, assign identities and policies, track performance and behavior, manage agent lifecycles and monitor costs. The system is intended to cover both Salesforce and third-party AI.
Many of the technologies underlying Enterprise AI Harness are available today. Salesforce plans to begin rolling out new capabilities and a unified interface in early 2027. The company has not yet announced final pricing and packaging.
