Agentforce

Chapter 13How Do You Test an Agent in a Sandbox?

In my younger and more impressionable years, I performed a task at a digital advertising agency called “measurement,” which is more interesting than it sounds. My task was to come up with a plan to determine the impact of an ad campaign – that is, how much it affected sales of the product. Campaign measurement was then and is now a complex alchemy of art and science and pushes data-driven marketing to its mathematical apogee.

And yet it was not respected. Especially during the era of flashy home page takeovers and Superbowl tie-ins, the creative idea always got 55 minutes of any 1-hour meeting. The media and measurement plans got 2 minutes each and were often skipped without a backward glance.

I tell you this to illustrate that the most important steps in a process are often not the most glamorous. Building an agent is fun (trust me – try it). But testing an agent to see if it works? More important.

From the beginning, Agent Builder was packaged with something called Plan Tracer, which was designed to make sure agents were transparent about what they were doing and why. It provides a kind of X-ray into Agent Builder so a human user can watch it at work.

Figure 13.1 shows what it looked like at launch.

On the right side, you can see a sample dialog that mimics a real-world conversation between an agent and a customer. In this example, based on the real case of using Agentforce agents to help attendees at Dreamforce plan their ...