On building an Agent with UiPath

Last month we got to attend a hackathon aimed at using the new agent builder feature in UiPath, which is currently in preview. As a group of people we got together to implement a shipment tracking agent in order to see how good such an agent would perform, how easy it is to build one and if we believe it to be a viable solution for the kind of use cases, that we deal with regularly. 

AI-powered control center tracking global cargo shipments by Midjourney

The use case is fairly straight forward. Imagine you are some kind of logistic company that receives emails from customers asking about their shipments. Some twenty years ago, we worked in customer service at an online retailer, so we were somewhat reminded of our time back then. The agent should understand the email, if there is a tracking number provided it should use it to call a databases API, understand the response from the db and write an email response accordingly. Obviously there is many ways to do this, but in our case we wanted to have an intelligent agent do it, and all it got was a prompt and an API plus its description. 

We hadn't actually been that familiar with UiPath, so going into the hackathon we were curious how far we'd get. To our delight, we were able to get an agent up and running within an hour that could perform the task as required! We described the output email the agent has to produce right in the prompt. For actual production we'd recommend using some kind of template. The agent understood how to use the API call tools without us describing them (merely from their names), but the UiPath people recommend adding the descriptions and emphasized that they're important. For cases where the agent can not process the input, eg cause something crucial is missing, the agent builder offers to implement escalations, which are forwarded to humans. Our favorite feature though was the testing page. One can implement test sets of input output pairs and metrics for evaluating them. This way one may even apply test driven development mindsets to agent building. 

Screenshot of UiPath Agent Builder (preview) with our system instruction, example prompt and the agents' result. We used a real example for the testing which is why we blacked out the information.

Lastly we want to compare UiPath's agents with some other solutions we've seen. Namely Microsofts Copilot Studio and Cognigy's agents. We've already blogged about Copilot Studio last year; but so far we've mainly looked at it as a chatbot solution. Recently we built a prototype of a Copilot Agent, i.e. a Copilot chatbot equipped with Power Automate cloudflows, which can be triggered on a users behalf. While we found it less intuitive than UiPath, it also worked well, once we figured things out.

One of the most intruiging chatbot solutions we've seen is Cognigy. On a side note, we will be attending a Cognigy conference soon, which we plan to blog about too. Unlike UiPath and Power Automate, Cognigy does not have any automations (cloudflows or RPA), and is focused on chats, conversations and calls. In order to build an agent with Cognigy, one can connect it to automations in either Power Automte or UiPath and give these automations as tools to the agent. Recently we were also involved in building an agent demonstrator using cognigy and we found it the most intuitive of the three. Especially when figuring out the thought process of the agent, Cognigy was most accessible.