Salesforce AI Research today introduced eVerse, an enterprise simulation framework built to train AI agents through synthetic data, stress-testing, and reinforcement learning from human feedback (RLHF). Announced at Dreamforce and detailed in new analyst materials, eVerse targets what Salesforce calls “jagged intelligence”—the tendency of large language models to ace complex reasoning yet stumble on simple, real-world tasks.
The framework runs through three phases: Synthesize (realistic enterprise environments via CRMArena-Pro), Measure (stress-testing agents across voice and text), and Train (closing performance gaps with RLHF from domain experts). Used to develop Agentforce Voice and now piloted at UCSF Health, eVerse has boosted enterprise task success from 19 to 88 percent, with roughly 70 percent retention and 60 percent generalization to new cases.
“With eVerse, we tested every nuance of human conversation before Agentforce Voice reached production,” said Silvio Savarese, Salesforce EVP and Chief Scientist. “That rigor turns breakthrough research into dependable customer experiences.”
The Post-Training Problem: Getting LLMs to Behave
Large language models trained on Internet-scale data still struggle to perform real business tasks reliably. A model that can pass the bar exam may fail to route a support call when background noise or edge cases appear.
Reinforcement learning—training agents in simulated environments that reward correct behavior—has emerged as the fix. OpenAI’s o1 and Anthropic’s Claude Opus 4 both used RL to surpass earlier reasoning benchmarks. As pre-training gains flatten, labs are investing billions into “RL gyms,” synthetic workplaces where agents learn through trial, error, and feedback.
Two rival architectures have since formed—overlay and embedded—and their divergence may define competitive advantage in the agentic-AI era.
From Robotics Simulation to Enterprise AI
Savarese traces eVerse’s lineage to robotics: “You can’t train robots in the real world without first simulating them,” he explained.
Before Salesforce, he led Stanford’s Computational Vision and Geometry Lab and the SAIL-Toyota Center for AI Research, focusing on computer vision and 3D perception. His spouse, Fei-Fei Li, co-creator of ImageNet and now CEO of World Labs, just launched Marble, a system that generates interactive 3D worlds from text or video—part of what she calls “spatial intelligence.”
Robotics researchers have long confronted the “reality gap”—the delta between simulated and real-world performance. Their solution: generate synthetic scenarios, measure failures, update simulations, and iterate.
“Feedback from real-world use is what closes that gap,” Savarese noted.
eVerse applies this robotics discipline to digital work. Instead of training robots to navigate warehouses, it trains AI agents to navigate CRM workflows, billing systems, and service interactions—closing the “reality gap” for enterprise data and processes.
Overlay vs. Embedded: The Great Enterprise AI Split
The overlay–embedded divide mirrors the Overlay–Embedded Framework we’ve tracked throughout 2025 (KeenanVision.net).
The Overlay Path: RL Gyms for UI Automation
OpenAI, Anthropic, and others train models to operate SaaS apps from the outside—simulated Salesforce, Zendesk, or Cerner interfaces—so agents can “click” through mock UIs. Each expert demonstration improves a single training run, creating a costly, human-bound feedback loop.
One OpenAI executive described the goal as turning “the entire economy into an RL machine.” But that economy still depends on experts recording every workflow.
The Embedded Path: Platform-Native Simulation
Salesforce’s eVerse takes the opposite route—training inside the platform. CRMArena-Pro generates synthetic “digital twins” with realistic data, workflows, metadata, and logic. Instead of watching what users see, agents interact with what Salesforce knows: Data Cloud records, metadata schemas, automation rules, and trust-layer policies.
Because 95 percent of enterprise orgs are customized, generic RL gyms cannot generalize. eVerse solves this by simulating each org’s configuration, enabling agents to learn native Salesforce behavior through the same APIs and data models that drive production systems.
Exponential Learning and the New Competitive Moats
This embedded architecture activates what Virtual Employee Economics defines as the Law of Exponential Learning: improvements propagate instantly across all agent instances through local experience, network sharing, and guided RLHF.
Salesforce’s “continuous loop” flywheel runs as follows:
- Synthesize – generate thousands of synthetic enterprise scenarios
- Measure – stress-test agents on realistic voice/text interactions
- Train – apply domain-expert RLHF (e.g., UCSF Health…
Source link
