Description Transcript
In this video, we dive deeper into a full buildout of a project manager agent that can read through meeting transcripts and automatically assign action items to the team to keep everybody on the same page.
To learn more, check out https://retool.com/agents
0:04 When you're on a team with a bunch of 0:05 different projects and priorities, it 0:06 can be a struggle to keep everyone 0:08 working together smoothly. You're 0:09 constantly bouncing back and forth 0:10 between Google Docs, meeting 0:12 transcripts, project management tools, 0:13 and more just to keep your team working 0:15 together. Luckily, we can build an AI 0:17 agent in Retool that can help us 0:19 automate this entire process. Let's take 0:20 a look at how that works. To get 0:22 started, we'll head over to our retool 0:24 dashboard where we'll see the agents 0:25 tab. Once we're here, we can click on 0:27 new agent and let's just start from 0:29 scratch. After we fill out some of the 0:31 details about our agent, we're brought 0:32 to the configuration screen where we can 0:34 prompt our agent via instructions. In 0:36 this case, we let our agent know that 0:37 it's a project management agent 0:39 supporting the applied AI team at 0:40 Retool. And its goal is to make sure 0:42 that everything it knows and all the 0:43 information that it observes is tracked 0:45 in Jira. And so, we tell it that it's 0:47 able to listen to call transcripts and 0:48 cross reference what it hears with items 0:50 in Jira, as well as create new Jira 0:52 tickets to support the team. We 0:54 mentioned that after it's completed the 0:55 work that it needs to do, we want it to 0:57 write a brief project update and send it 0:58 out via email. And we give it a couple 1:00 more style guidelines around how it 1:02 should respond and how it should behave. 1:05 After that, we can come down here and 1:06 pick our model from a variety of 1:07 providers, but let's choose GPT 4.1 for 1:10 now. If we need to configure the 1:12 advanced settings such as temperature or 1:14 the maximum number of iterations that 1:16 the agent can perform through its loop, 1:17 we can do that here as well. Then we 1:20 need to configure tools. In this case, 1:22 we'll be creating a couple custom tools 1:24 as well as using some of the pre-built 1:25 tools that are already provided in 1:26 retool. Let's start by choosing the get 1:28 Google Doc and search Google Docs tool 1:30 so that our AI agent can browse through 1:32 our Google Drive and search for our 1:33 meeting notes. Then, we're going to need 1:35 to create a couple custom tools. If we 1:37 click create new custom tool, you can 1:39 see we can provide a name and 1:40 description and even optionally require 1:42 user confirmation before those tools 1:44 used. In this case, we'll jump into the 1:46 function editor and create a custom tool 1:49 that pulls our meeting transcripts from 1:50 the Zoom REST API. This means that when 1:53 our meeting is concluded, we can 1:54 automatically pull our transcripts into 1:56 our agent. In addition, we'll also build 1:58 a custom tool that lets us create Jira 2:00 tasks using Retool's native Jira 2:02 integration. We can create a new block 2:04 for a resource query and select the Jira 2:07 integration from the drop down. Then, we 2:09 can fill out all the relevant details to 2:11 give our agent access to create tickets 2:13 in Jira. It's important when you're 2:15 creating tools to make sure all the 2:16 parameters are filled out correctly with 2:18 proper descriptions. This allows our AI 2:20 agent to decide between which tools to 2:21 use and make sure that the tools that 2:23 we've created have the correct 2:25 parameters passed to them. Now that we 2:27 have our agent configured and our tools 2:28 created, let's check out the chat tab. 2:30 In this tab, you can start a chat with 2:32 your agent and you can watch it respond 2:33 to your messages, call tools, and take 2:35 action in real time. This is one of the 2:38 main ways that users will interact with 2:39 agents in Retool. But in this case, 2:41 because our standup meeting runs on a 2:42 given schedule, we want to actually 2:44 trigger our agent via a workflow. So, if 2:47 we head over to the workflows editor, we 2:48 can create a new workflow to trigger our 2:50 agent on a schedule. In this case, let's 2:52 say our agent needs to run at 2 p.m. 2:54 Central time every week because our 2:56 standup meeting is the hour before. We 2:58 can set that up via a trigger, and we 3:00 can create a workflow block that calls 3:02 our agent and kicks off its run. Now 3:04 that we have our agent configured and 3:06 being triggered via workflow, let's take 3:07 a look at our eval tab. We can create a 3:10 new data set on which to run our eval 3:12 from here. And in this data set, what we 3:15 want to do is make sure that our agent 3:16 is calling the correct tools. So we'll 3:18 add a test case that when the user 3:20 mentions sending an email, it calls our 3:22 email tool. We can fill out our input, 3:25 select tool choice as the type of test 3:27 case, and then choose the expected tool 3:29 that we want the agent to call. Once we 3:32 fill in some information about the 3:33 expected parameters, we can hit create, 3:35 and we see that our test case is set up 3:37 here. After filling in a few more test 3:39 cases, we can tab over to 3:42 eval set. After a few seconds of our 3:45 eval running, you'll see whether our 3:46 eval passed or failed, and you'll be 3:48 able to dive in deeper on what happened 3:50 here and see what's going on with your 3:51 agent. This is a great way to make sure 3:53 your agent is running smoothly, even if 3:55 you switch models, update the prompt, or 3:57 add new tools. And as our agent starts 3:59 to run, the logs tab will be populated 4:01 with all of the actions that it's taking 4:02 and all of the thoughts that it's 4:03 having. This is a great place to go once 4:06 our workflow starts triggering our agent 4:07 to run as we can inspect all of the past 4:10 runs and thought processes that our 4:11 agent is having here. And just like 4:13 that, our project manager agent can read 4:15 through our meeting transcripts, access 4:16 any of our notes, and assign action 4:18 items to team members automatically. 4:20 When you're ready to build an AI agent 4:21 just like this, check out Retool and 4:23 build your first agent today. 4:26 [Music] 4:27 [Applause]
