Salesforce previously enabled its customers to build AI agents through Agentforce and was far ahead of the curve in doing so. The company is going the extra mile and now also offers the logical next step, a set of tools to test, monitor, and tweak such agents, under the umbrella of Agentic Lifecycle Management.
The new toolkit should enable companies to prototype, scale and optimize AI agents securely without disrupting live production environments.
We’ve written about AI agents, essentially customized mini-AI models that often focus on one specific task; see Exhibits A, B, and C. As more offerings become available in this area and more companies show interest in them, the key is to ensure that these agents also (continue to) work well.
Often, their training happens with proprietary data or depends on sensitive information for their training and operation. Furthermore, there are all kinds of operational hurdles to overcome, and the risk of incorrect input always remains. The Agentic Lifecycle Management package tools allow further testing, monitoring, and adjustment via no-code applications.
A ‘limitless workforce’
“Agentforce Agentforce is helping businesses create a limitless workforce,” said Adam Evans, who is at the helm of the AI Platform at Salesforce. “To deliver this value fast, CIOs need new tools for testing and monitoring agentic systems.”
That happens from the new central Agentforce Testing Center application, which allows admins to simulate hundreds of customer interactions using AI-generated test cases. This allows companies to refine the behavior of their agents, which should improve their ability to deal with real-world scenarios. For example, it is possible to run similar cases in parallel and see how often the agent gives the desired answer. Then, the agent can be instructed to give the most desirable or useful answer more often in appropriate cases.
Training before the game
Once the model is well underway, it is time to subject it to all kinds of stress tests in Sandboxes. These are isolated training fields that mimic production environments as closely as possible. That makes it possible to test them without affecting the live environment. When their training is done, the agents can be transferred to participate in ‘the game’ i.e. live environments. This functionality is also available for Data Cloud, the foundation of Salesforce’s AI agents.
Tip: Marc Benioff is on fire, never before has the Salesforce CEO been so outspoken
In addition to these options, Salesforce also offers advanced monitoring in this new package. The Agentforce Analytics and Utterance Analysis tools provide detailed insights into agent usage and performance. These solutions are interwoven with the Einstein Trust Layer. This ensures that the AI agent’s action space remains compliant with company policies and can securely draw on existing, proprietary data sources such as product details, purchase history, customer preferences, warranty and inventory information. This closes the agent feedback loop.
Of course, all this processing, testing, and honing of agents costs money. Salesforce’s Digital Wallet provides a detailed overview of the resources used and allows users to set up alerts if resource consumption threatens to become excessive.
Also read: Salesforce’s Agentforce reaches general availability with sky-high expectations