Take a look at what’s inside Agentforce 3

Salesforce this week announced Agentforce 3, which the company says will give Agentforce customers the visibility and control they need to scale their AI agents.

Visibility is a fundamental problem for early AI agent deployments. Simply put, teams can’t see what their agents are doing. It’s no surprise Agentforce 3 takes steps to increase control for businesses deploying AI agents.

Here’s a look at the key elements of Agentforce 3.

Agentforce Command Center

Agentforce Command Center, which is built into Agentforce Studio, is an observability solution that delivers a single pane of glass to monitor agent health, measure performance and optimize outcomes. 

Teams can use Command Center to analyze every AI agent interaction, drill into specific moments, understand usage trends, and see AI-powered recommendations for tagged conversation types. This can all be used to continuously improve the agent experience. 

Live, detailed analytics will help teams monitor for latency and escalation frequency, while real-time alerts can be triggered when the unexpected happens.

Command Center offers detailed dashboards that track agent adoption, feedback, success rates, cost and topic performance, which means teams can see what’s gaining traction and where to improve.

Agentforce captures agent activity in a native, extensible session-tracing data model in Salesforce Data Cloud, which powers analytics, monitoring and real-time alerting. The OpenTelemetry standard allows teams to integrate these agent signals with popular tools, including Datadog, Splunk, Wayfound and other Salesforce monitoring partners.

Dig deeper: Salesforce Agentforce: What you need to know

Model context protocol (MCP) support

Agentforce will soon include a native model context protocol (MCP) client, allowing Agentforce agents to connect to any MCP-compliant server without any custom code. That gives access to enterprise tools, prompts and resources — governed by existing security policies.

Central to the MCP strategy are a pair of companies owned by Salesforce.

New MCP connectors for MuleSoft will convert any API and integration into an agent-ready asset, complete with security policies, activity tracing and traffic controls.

Heroku Managed Inference and AppLink will make it fast and easy to deploy, register, maintain and connect custom MCP servers. With Heroku’s secure infrastructure and DevOps automation, developers can bring trusted custom actions to Agentforce.

Agentforce customers can discover MCP servers from more than 30 partners through AgentExchange. Launch MCP partners include AWS, Box, Cisco, Google Cloud, IBM, Notion, PayPal, Stripe, Teradata and WRITER.

Dig deeper: Salesforce introduces new Agentforce pricing models

Enhancements to the Agentforce architecture

Agentforce 3 includes an enhanced Atlas architecture with lower latency, greater accuracy, global availability, and additional control options through new LLMs hosted on Salesforce infrastructure.

Agentforce can now use Anthropic’s Claude Sonnet model hosted via Amazon Bedrock within the Salesforce trust boundary to meet customers’ needs in high-compliance industries. Anthropic will work with Salesforce to help customers in regulated industries scale Agentforce adoption with Claude. Later in 2025, Salesforce will also allow customers to use Google’s Gemini in Agentforce.

Other Agentforce 3 enhancements include:

• Trusted, expanded grounding by using web search as a data source, allowing agents to go beyond internal data to answer requests and deliver inline citations referencing the sources used in responses.
• An expanded global footprint, deploying to Canada, the U.K., India, Japan and Brazil to serve AI agent traffic within those regions. 
• Support for six new languages, including French, Italian, German, Spanish, Japanese and Portuguese. Salesforce expects more than 30 additional languages to roll out in the coming months.
• Automatic, latency-based failover — dynamically shifting traffic between model providers in case of performance degradation or outages.