Salesforce Launches eVerse: An Agent Simulation Platform for Text and Voice

Salesforce has just announced the launch of eVerse – what its AI research arm is calling “an enterprise simulation environment for voice and text agent training”. 

eVerse, built to test agents using lifelike, simulated conversations and scenarios, does this through three main methods: synthetic data generation, stress-testing, and reinforcement learning. 

What Is eVerse?

eVerse forms part of the next generation of agent testing – testing that needs to feel more relevant, applicable, and adaptable. 

At present, agents and artificial intelligence in general have the capabilities to solve PhD-level problems, but struggle with basic tasks, especially if the input from a human doesn’t match what it is expecting or has been trained on. 

Salesforce uses the example of the riddle “Where does Christmas come before Thanksgiving?”. One of the market’s top LLMs is likely to answer with “in the dictionary” – because alphabetically, ‘C’ precedes ‘T.’ That nuance, context, and wider understanding are not yet there, and this is what is being referred to as “jagged intelligence”. 

This is where eVerse comes in. Businesses can now test their agents through a chronological framework that focuses on three stages: synthesize, measure, and train.

Synthesize

Synthesizing is the first step because, logically, the equation here is simple: in order to get the best agents, you need to train them in the best, most realistic environment. 

The foundational work for eVerse was established with CRMArena-Pro, an agent training environment that Salesforce introduced earlier this year. Here, the training grounds are too completely synthetic, allowing agents to interact with realistic customer data, multi-step workflows, and the edge cases that make business operations unpredictable.

Through this work, Salesforce was able to confirm that 90% of domain experts rate its synthetic data generation as realistic or very realistic. 

Now, eVerse takes this a step further. One of its strongest capabilities is its synthetic voice simulations, designed to train AI to deal with unpredictable customer interactions. It is all well and good that an agent is able to help a calm, slow-talking customer, but what happens if their connection is spotty, there is lots of background noise, or they’re also half-speaking to someone else while on the call?

With eVerse, you can create specific Personas – say a distracted dad or a frustrated caller – and save them for recurring use. Making Personas appears to be fairly simple, too, with the option to select multiple differentiators, such as audio that continuously cuts out or a thick accent.

Measure 

Measuring is the next step. For each persona and variable, you can perform numerous tests. eVerse features an ‘AI Judge’ that will give you specific feedback for each test, and this will form part of the decision as to whether or not the agent passes or fails the test.

“[You] can actually run many different sessions, hundreds, or even thousands of sessions,” Itai Asseo, the Head of Incubation & Brand Strategy at Salesforce AI Research, explained in a media preview call. 

If your agent does fail a test, eVerse explains the exact reason why, whether that be the prompt, the action, or the topic. 

Train

Once the vulnerabilities and gaps are identified, eVerse’s training engine closes them through reinforcement learning guided by human expertise. Not only will the testing data show how different test runs improved the agent over time, but a human can guide its training further with detailed information on its accuracy, latency, and more.

Salesforce claims that its research using this method demonstrated significant improvements – specifically, doubling the performance rate on enterprise tasks (from 19% to 88% success rates).

Agentforce Voice Case Study

Salesforce has also revealed that eVerse has been crucial to the development of Agentforce Voice, its latest voice capabilities installment to its proprietary AI. 

This has proved to be more important than ever, because although it can be assumed most users looking to get a problem solved through a customer service route are comfortable typing and messaging, calling a help line is still a popular go-to channel.

eVerse is the platform that allows Agentforce Voice to effectively manage even the most unusual voice conversations. Through eVerse’s virtual environments, which accurately replicate the noise, accents, crosstalk, and complexity of real-world operations, AI agents are trained and stress-tested.

Before its launch, Agentforce Voice underwent thousands of simulated conversations in eVerse. This rigorous testing enabled teams to identify and resolve errors early, ensuring the enterprise-grade reliability that customers require.

“With eVerse, we were able to explore many nuances of human conversation before Agentforce Voice reached production,” said Madhav Thattai, COO of Salesforce AI. “This type of rigor is what turns breakthrough research into scalable products and dependable customer experiences. It’s how we’re expanding that same level of responsiveness and consistency across the full observability stack to solve our customers’ most complex needs.”

What Does It Cost?

In terms of costs, Salesforce has said that this is currently being finalized. That part of the ongoing testing with eVerse is to establish how many “tokens” – the currency used for each test –  are to be consumed and when. 

“For clarity, today, customers either buy our agentic capabilities as a part of their licences with our Agentforce License products, or they buy them as a monetization that we do with actions – which is how many actions the agent is actually taking in production,” Madhav explained. “Any customer that uses Agentforce gets the platform tools. What gets monetized are the actions that they actually deliver in their production environments.”

As of today, eVerse is being rolled out in a pilot to select Salesforce customers. 

Final Thoughts

eVerse officially marks Salesforce’s next advancement into AI agent creation and testing, opening up the capabilities for businesses to build fully-dimensional, intelligent, and nuanced agents. 

Although there are likely to be challenges – including further questions around pricing, security concerns, and closing any kind of “reality gap” between what Salesforce promises and what it actually delivers – it will be interesting to see how eVerse performs and what kind of agents it has the potential to create.