Salesforce adds Testing Center to Agentforce for AI agents

Salesforce has added a new set of tools under the name of Testing Center to its agentic AI offering, Agentforce, to help enterprise users test and observe agents before deploying them in production.

Agentforce, which was made generally available in September, is a low-code suite that enables enterprises to build AI agents that can reason for themselves when completing sales, service, marketing, and commerce tasks.

The autonomous nature of these agents is a central facet of agentic AI, a rising enterprise strategy for transforming business processes by automating specific functions within those processes, without human intervention.

The tools added as part of the Testing Center upgrade include generating synthetic interactions using natural language interactions, sandboxes, and tools for observing the agents’ performance.

AI-generated testing and sandboxes

The ability to generate synthetic interactions, according to Salesforce, will come in handy as it will allow enterprise users to test all the different ways a customer may interact with an agent and check if the agent is performing as required.  

One of the key tests to check if the agent is performing as required is to determine if the agent selects the right topic and action at scale basis the given input.

In case, enterprise users see that the agent is not performing as required, they can use the test data to refine instructions to the agent to ensure that it selects the right topic and action.

This capability is over and above the Plan Tracer feature that comes packaged inside Agentforce. Plan Tracer helps enterprise users investigate the reasoning process of an agent and make tweaks via the Agent Builder module if necessary.

In addition, Salesforce said that it was making Sandboxes for Agentforce and Data Cloud generally available in order to help enterprises test agents in an isolated and safe environment.

Sandboxes, according to Salesforce, work by mirroring images of an enterprise’s production data and configurations.

“By replicating the enterprise’s data and metadata into a risk-free environment, development teams can rapidly assemble their unstructured data foundation and rigorously prototype Agentforce without fear of disrupting the business,” the company explained in a statement.  

From testing to production

Once enterprise users determine that the agents are functioning as required, they would be able to deploy the changes using proprietary tools such as Change Sets, DevOps Center, and the Salesforce command line interface (CLI), the company said.

Separately, the company said that the Sandboxes will allow enterprise teams to test the Einstein Trust Layer — the AI guardrails offering — before putting the agents into production.

“With the Einstein Trust Layer’s audit trail and feedback store in sandboxes, teams can build a closed loop for AI testing — iterating on prompts and actions based on user feedback,” the company said, adding that once Agentforce is live in production, new capabilities for granular insights into agents become available through Agentforce Analytics and Utterance Analysis.

These new observability tools can be used for continuous iteration of agents. The company said the usage of sandboxes and other Testing Center tools can be checked via the company’s Digital Wallet offering, which is an account management tool.