Salesforce Agentforce 3: Building the Production Infrastructure for Enterprise AI Agents
Salesforce’s rapid evolution of Agentforce has exposed a critical gap in the enterprise AI landscape. While companies have rushed to deploy AI agents, they’ve lacked the fundamental tooling to operationalize them at scale. The absence of robust observability and management capabilities has left enterprises flying blind, unable to see what their agents are doing or optimize their performance.
With Agentforce 3, Salesforce is attempting to close this gap by delivering what Salesforce EVP and GM of Salesforce AI Adam Evans describes as the infrastructure for “bringing agents closer into cross-business use cases in a hybrid workforce” – essentially productionalizing what has largely been experimental deployments.
Table of contents
What Was Announced
Several things were announced today, teeing up a week of Agentforce announcements and webinars.
Agentforce Command Center: Enterprise Observability for AI Agents
The centerpiece of Agentforce 3 is the Command Center, a complete observability solution that gives leaders a unified pane of glass to monitor agent health, measure performance, and optimize outcomes. Built directly into Agentforce Studio, Command Center represents a significant maturation of the platform.
Key capabilities include:
- Interaction Analysis: Teams can analyze every AI agent interaction, drill into specific moments, understand trends in usage, and see AI-powered recommendations for tagged conversation types
- Real-time Health Monitoring: Live, detailed analytics for latency, escalation frequency, and error rates, plus real-time alerts when the unexpected happens
- Performance Dashboards: Detailed dashboards that track agent adoption, feedback, success rates, cost, and topic performance
- OpenTelemetry Support: Built on the OpenTelemetry standard, these agent signals integrate seamlessly with tools teams already use, including Datadog, Splunk, Wayfound, and other monitoring partners
The addition of OpenTelemetry support deserves particular praise. By adopting this open standard, Salesforce acknowledges that enterprise AI monitoring cannot exist in isolation. Organizations need their AI agent telemetry to flow into existing observability stacks where they can correlate agent behavior with broader system performance. This represents a mature approach to enterprise infrastructure that has been sorely missing from the AI agent space.
Model Context Protocol (MCP): Opening the Connectivity Floodgates
Perhaps the most strategically significant announcement is native MCP support within Agentforce, enabling agents to connect to any MCP-compliant server without custom code. Salesforce is positioning MCP as the “USB-C for AI,” and the implications are profound.
The MCP implementation includes:
- Native Client Support: Agentforce will include a native MCP client, enabling access to enterprise tools, prompts, and resources — governed by existing security policies
- MuleSoft Integration: MuleSoft converts any API and integration into an agent-ready asset, complete with security policies, activity tracing, and traffic controls
- Heroku Hosting: Heroku Managed Inference and AppLink make it fast and easy to deploy, register, maintain, and connect custom MCP servers
The expanded AgentExchange now features MCP servers from 30+ partners including AWS, Box, Cisco, Google Cloud, IBM, Notion, PayPal, Stripe, Teradata, and WRITER. This ecosystem approach begins to blur the lines between Salesforce’s traditionally embedded AI architecture and the characteristics of overlay AI vendors.
Enhanced Atlas Architecture: Trust at Scale
Agentforce 3 features an enhanced Atlas architecture with 50% lower latency since January 2025. Notable improvements include:
- Expanded LLM Choice: Agentforce can now use Anthropic’s Claude Sonnet model hosted via Amazon Bedrock within the Salesforce trust boundary, with Google Gemini support coming later this year
- Web Search and Citations: Agents can now go beyond internal data to answer requests, with inline citations that provide references to grounding sources
- Global Expansion: Deploying to Canada, the U.K., India, Japan, and Brazil with support for six new languages
- Automatic Failover: Dynamic shifting of traffic between model providers in case of performance degradation or outages
Industry-Specific Acceleration
Agentforce 3 introduces more than 200 pre-built industry actions, half of them new this summer, from patient scheduling to advertising proposal generation, to vehicle servicing. Combined with new flexible pricing including Agentforce add-ons starting at $125 per user per month with unlimited employee agent usage, Salesforce is clearly targeting rapid enterprise adoption.
The MCP Factor: Blending Embedded and Overlay AI Architectures
The introduction of native MCP support represents a strategic shift for Salesforce. Traditionally, Salesforce has epitomized the embedded AI approach – AI capabilities tightly integrated within the platform’s data model and workflows. MCP support begins to incorporate overlay AI characteristics, where agents can interact with external systems more fluidly.
As Gary Lerhaupt, VP, Product Architecture of Agentforce, explained during the launch, MCP provides “a universal translator for external tools” but emphasized that “MCP is great, especially when combined with security, governance and trust”. This balance between openness and control is critical – Salesforce is maintaining its enterprise-grade security posture while enabling the interoperability that production deployments demand.
Reality Check: Enterprise AI at Scale Remains Elusive
Despite these advances, a sobering reality persists. CIOs across the industry remain largely unsatisfied with any vendor’s ability to deliver Enterprise AI success at scale. Neither Microsoft with Copilot, ServiceNow with Now Assist, nor any other major embedded AI vendor has demonstrated repeatable, measurable success in production deployments that justify their astronomical price tags.
AI orchestration remains more art than science. The industry is collectively working through fundamental questions about agent reliability, governance, and ROI measurement. While Salesforce’s Command Center provides better visibility than competitors, it’s still early days for understanding how to optimize multi-agent systems operating at enterprise scale.
What Salesforce is doing well is listening to its early adopters and rapidly iterating. As Adam Evans noted, the platform has evolved from basic uptime monitoring to task-level performance tracking, and now to business KPI alignment where “metrics mirror more towards higher level business metrics”. This evolution reflects the real journey enterprises are taking as they move from pilots to production.
Building the Agentic Layer
The good news is that Salesforce is steadily assembling the components its customers need to create what amounts to an “agentic layer” around the enterprise. Command Center provides observability. MCP enables connectivity. The enhanced Atlas architecture delivers reliability. Pre-built industry actions accelerate deployment.
Early results are encouraging: Engine reduced average customer case handle time by 15%, 1-800Accountant autonomously resolved 70% of administrative chat engagements during tax season, and Grupo Globo increased subscriber retention by 22%. These metrics, while impressive, represent early adopters with focused use cases rather than the wall-to-wall enterprise deployments that will define success at scale.
Looking Ahead
Agentforce 3 represents necessary table stakes for enterprise AI agent deployment rather than a revolutionary leap. The addition of proper observability through Command Center and OpenTelemetry support brings Salesforce closer to parity with traditional enterprise software expectations. MCP support acknowledges that agents must operate beyond platform boundaries.
The real test will come as enterprises attempt to scale beyond departmental deployments to true hybrid workforces where hundreds or thousands of agents collaborate with human workers. The infrastructure Salesforce is building with Agentforce 3 provides the foundation, but the industry collectively has much to learn about making this vision reality.
For now, CIOs should view Agentforce 3 as evidence that Salesforce is committed to the long journey of making AI agents enterprise ready. The platform is becoming more observable, more interoperable, and more reliable. Whether it becomes truly manageable at scale remains to be seen.
