Salesforce Agentforce 3 aims to tackle what many businesses have been struggling with: actually seeing what their AI agents are up to.
Since its debut back in October 2024, Agentforce has been racking up some wins across a variety of sectors. Engine managed to slash customer case handling times by 15 percent, while 1-800Accountant handed off 70 percent of administrative chat queries to AI during the madness of tax season.
But what’s interesting about this upgrade isn’t just the numbers, it’s how Salesforce is addressing the elephant in the room that nobody likes to talk about: businesses are deploying AI agents at breakneck speed without really understanding what they’re doing or how to improve them.
Keeping tabs on your agents
The centrepiece of Agentforce 3 is what Salesforce calls the Command Center (essentially a mission control for your AI employees.) It lets managers peek under the bonnet to spot patterns in how agents are performing, track health metrics in real-time (latency, escalation rates, errors), and identify which bits are working versus which need a swift kick.
For anyone who’s ever deployed AI tools and then wondered “now what?” this level of visibility could be game-changing. The system captures all agent activity using the OpenTelemetry standard, which means it plays nicely with tools like Datadog and Splunk that your IT team probably already has on their screens.
AI adoption is absolutely skyrocketing. Forthcoming data from the Slack Workflow Index shows AI agent usage up 233 percent in just six months. During that time, about 8,000 organisations signed up to deploy Agentforce.
Ryan Teeples, CTO at 1-800Accountant, said: “Agentforce autonomously resolved 70% of 1-800Accountant’s administrative chat engagements during the peak of this past tax season, an incredible lift during one of our busiest periods. But that early success was just the beginning.
“We’ve established a strong deployment foundation and weekly are focused on launching new agentic experiences and AI automations through Agentforce’s newest capabilities. With a high level of observability, we can see what’s working, optimise in real time, and scale support with confidence.”
Salesforce Agentforce 3 doesn’t just provide data, it actually suggests improvements. The AI effectively watches itself, identifying conversation patterns and recommending tweaks. It’s a bit meta, but potentially very useful for overstretched teams who don’t have time to manually review thousands of bot interactions.
The connectivity conundrum solved?
Another headache Salesforce is tackling is connectivity. AI agents are only as useful as the systems they can access, but connecting them securely to your business tools has been a pain for most organisations.
Agentforce 3 brings native support for Model Context Protocol (MCP) – which Salesforce rather aptly describes as “USB-C for AI.” This essentially means AI agents can plug into any MCP-compliant server without custom coding, while still respecting your security policies.
This is where MuleSoft (which Salesforce acquired a few years back) comes into play, converting APIs and integrations into agent-ready assets. Heroku then handles deployment and maintenance of custom MCP servers.
Mollie Bodensteiner, SVP of Operations at Engine, commented: “Salesforce’s open ecosystem approach, especially through its native support for open standards like MCP, will be instrumental in helping us scale our use of AI agents with full confidence.
“We’ll be able to securely connect agents to the enterprise systems we rely on without custom code or compromising governance. That level of interoperability has given us the flexibility to accelerate adoption while staying in complete control of how agents operate within our environment.”
Growing the Salesforce Agentforce ecosystem
Perhaps the most interesting aspect of this announcement isn’t what Salesforce built themselves, but the ecosystem they’re nurturing. Over 30 partners have created MCP servers that integrate with Agentforce, including players like AWS, Google Cloud, Box, PayPal, and Stripe.
These integrations go far beyond simple data access. For instance, AWS integration lets agents analyse documents, extract information from images, transcribe audio recordings, and even identify important moments in videos. Google Cloud connections tie into Maps, databases, and AI models like Veo and Imagen.
Healthcare appears to be a particularly promising sector.
Tyler Bauer, VP for System Ambulatory Operations at UChicago Medicine, explains: “AI tools in healthcare must be adaptable to the complex and highly individualised needs of both patients and care teams.
“We need to support that goal by automating routine interactions in our patient access center that involve common questions and requests, which would free up the team’s time to focus on sensitive, more involved, or complex needs.”
The real question, of course, is whether all this will actually help businesses manage the growing army of AI agents they’re deploying. Getting visibility into AI performance has been a blind spot for many organisations—they often know roughly what percentage of queries the AI is handling, but struggle to identify specific shortcomings or improvement opportunities.
Adam Evans, EVP & GM of Salesforce AI, says: “Agentforce 3 will redefine how humans and AI agents work together—driving breakthrough levels of productivity, efficiency, and business transformation.”
Whether it lives up to that lofty promise remains to be seen, but addressing the visibility and control gap is certainly a step in the right direction for businesses struggling to properly manage their AI initiatives.
