Description Transcript
Tune in to see Retool CEO David Hsu unveil Agents, designed for developers to automate work and unlock the ROI on AI.
This new class of software safely automates reasoning at scale, exponentially expanding what kinds of problems you can actually solve with software—today. We introduce the critical need for the app layer for AI and how industry leaders are already automating 100+ millions hours of work (and you can, too).
Explore Agents: https://retool.com/agents?utm_source=youtube&utm_medium=organic&utm_campaign=ai_launch&rcid=701Qo00000uFFa4IAG
Read more 0:03 Hi, I'm David from Retool. Large 0:07 language models are really smart. They 0:09 chat, reason, and help you code. LMS can 0:12 talk, but they can't act. To act, LM 0:16 need tools. That's why we've introduced 0:19 tool agents. We give LMS the tools to 0:22 execute real work inside your business. 0:25 Let's go see how it works. 0:28 On your left side, you see a typical 0:30 office workflow. And on the right side, 0:32 in the virtual office, you see agents 0:34 doing the same kinds of tasks, coding, 0:37 messaging, resolving issues. Much of 0:40 this work is repetitive and can be 0:41 automated, but only if LMS have 0:44 powerful, specific, and customized 0:46 tools. That's the key. So, let's zoom 0:50 in. On the left, our accountant 0:52 performing a multi-step manual process 0:54 to fight a chargeback in Stripe. takes 0:57 five minutes. So long, we had to in fact 1:00 speed it up by 10 times. And on the 1:03 right, retail agent number 17 is doing 1:05 the same task, but in real time. The 1:08 secret powerful predefined tools we gave 1:12 the LLM to go get chargeback from Stripe 1:14 to go gather evidence from your 1:16 Postgress database and submit it back to 1:18 Stripe. In those 5 minutes, agent number 1:21 17 finds 50 chargebacks. 1:24 Wow, so much 1:26 faster. That's the power of LM with the 1:29 right tools. And with Retool, you can 1:32 manage your agent and replay all the 1:33 work he's doing as if you were watching 1:36 over his shoulder. Or you can build Reut 1:38 agents to automate work all across your 1:41 company. create one to handle daily 1:43 project management tasks by listening to 1:45 standups. creating and assigning tasks 1:47 in Jira, following up on blockers, 1:50 syncing updates across different data 1:51 sources, and keeping everyone aligned 1:53 without lifting a finger, or another 1:56 agent to prepare materials for sales 1:58 calls by researching attendees on 2:00 LinkedIn, checking product usage from 2:02 your internal databases, pulling 2:04 internal notes from Salesforce, drafting 2:06 talking points, and generating a 2:08 personalized pitch deck in Google Slides 2:10 in just 2:12 minutes. or one to act as an executive 2:15 assistant to find time across packed 2:17 calendars, coordinating across different 2:19 time zones, rescheduling conflicts, 2:21 booking meetings with full context, and 2:23 assuring your calendar is always 2:25 perfect. With Retool agents, LMS finally 2:28 have specific custom tools to tackle 2:32 your business problems. Join the 2:35 thousands of other companies using 2:36 Retool to automate real work today. Our 2:40 customers have already automated over 2:43 100 million hours of labor using AI in 2:46 retool. That's a whole 5,000 personsized 2:49 company working for an entire decade. 2:53 That's nearly 5 billion in value. But 2:56 with rutual agents, we're setting an 2:58 even more ambitious goal. Automate 10% 3:01 of US labor by 2030. Rutual agents are 3:04 available now. Can't wait to see what 3:07 you build. 3:11 [Music] 3:18 You've just seen a glimpse of what we 3:19 think software could look like, not in 3:21 some distant future, but right now, our 3:24 goal is ambitious, but straightforward. 3:28 We are aiming to automate 10% of US 3:30 labor with retail agents by 2030. 3:34 Now, that might sound a little crazy, 3:37 but our customers have already automated 3:39 over a 100 million hours of real work 3:42 with Retool. To put that into 3:44 perspective, that's around 5 billion of 3:47 actual concrete value unlocked. That's a 3:51 whole 5,000 personsized company working 3:54 for an entire 3:56 decade. AI has had a weird few years. 3:59 Companies have poured billions of 4:01 dollars into AI, but largely speaking, 4:03 we've got chatbots. They're impressive, 4:05 but limited. They talk, but they don't 4:08 add. They don't actually integrate into 4:10 your business systems, trigger your 4:12 workflows, or perform real tasks. MCP is 4:15 great, but it's just getting started 4:17 right now. Chatbots haven't 4:18 fundamentally automated work yet. And as 4:21 a result, we have this enormous 4:23 disconnect between investment and 4:24 outcomes. In fact, there's roughly a 4:27 trillion dollar gap between we've put 4:29 into AI and we've gotten out of it so 4:31 far. So, what's missing? ALM can 4:35 conclusively pass the touring test, but 4:37 the problem is that they just can't do 4:39 anything yet. Even with MCP, which we 4:41 love and support with our product, many 4:43 of the tools just aren't specific enough 4:45 to your business. And so, if we're 4:47 serious about using AI to fundamentally 4:49 reshape productivity, we need more than 4:51 just sophisticated conversations. We 4:53 need an app layer for AI. A practical, 4:57 reliable, and scalable way to convert 4:59 powerful LLMs into software that does 5:01 actual work into software that solves 5:04 real business problems. This app layer 5:07 would bridge the gap between what an LM 5:09 can do and what a business needs done. 5:12 It's the layer that transforms AI from a 5:14 promising technology into measurable, 5:17 actionable productivity. 5:20 Now, some of you have already started 5:21 building AIdriven apps and retool. 5:23 That's great. But what we're introducing 5:26 today takes it further. Agents represent 5:29 a completely new type of software. They 5:31 automate reasoning at scale, safely, 5:34 unpredictably. They meaningfully expand 5:36 the scope of what software can do by 5:38 giving LMS real tools connected to your 5:41 actual data and 5:43 processes. It turns out there's a pretty 5:45 simple formula here. If you take 5:47 powerful LLMs, pair them with hypersp 5:49 specific, carefully designed tools, and 5:52 deploy them on retool, you get agents 5:54 that deliver meaningful impact to your 5:56 organization. Every retool query or 5:59 workflow you've written can now be used 6:01 by an 6:02 LLM. That is the secret to unlocking 6:05 value from LMS. LMS that don't just 6:07 chat, but actually 6:09 do. This isn't an incremental update. 6:12 This is Retool's next major chapter. 6:15 Today's launch marks the beginning of 6:17 our path towards a fully AI native 6:20 engine. A secure, reliable, 6:23 enterprisegrade and developer friendly 6:25 platform that allows you to leverage AI 6:28 everywhere across your company. And the 6:30 reason we're leading this effort is 6:32 because we've already built a 6:33 foundation. RTOL already powers mission 6:35 critical software at thousands of 6:37 companies. We're already the app layer, 6:39 the connective tissue between your data, 6:41 your logic, and your workflows. Every 6:44 retool primitive you've built is now a 6:46 tool an LLM can use. 6:49 Wow. And this is why you, the builders, 6:52 creators, and engineers are essential. 6:55 You are the ones who will take AI from a 6:57 series of impressive demos and transform 6:59 it into something tangible, powerful, 7:01 and actually useful. The real story here 7:04 isn't about AI. It's about what you can 7:06 actually achieve with 7:07 it. So, let's dive in. Kent will now 7:11 show you exactly how to build your first 7:13 retool agent right now and demonstrate 7:16 this is already changing what software 7:18 can accomplish. Hey all, I'm 7:21 Kent. Let's learn more about retool 7:24 agents. At its core, an agent is simple. 7:28 You give an LLM some input, access to 7:32 tools, and you let it run until the task 7:34 is done. 7:37 Although agents are simple, there are an 7:39 enormous number of decisions to make if 7:42 you're building agents from 7:43 scratch. Everything from the framework, 7:46 the cognitive architecture, providing a 7:48 cohesive model layer to the more 7:51 traditional considerations of 7:53 deployment, infrastructure, and 7:57 scalability. We saw our customers 8:00 rebuilding this AI orchestration layer 8:03 over and over. So we built retail agents 8:07 to take care of the undifferiated work 8:10 and give you a flexible powerful system 8:14 to automate real business processes 8:16 without reinventing the wheel every 8:19 time. So how does this work? Let's have 8:22 a closer look at the project manager 8:24 agent that we saw in the previous 8:27 clip. As the description says, this 8:30 agent helps manage projects for an 8:33 engineering team. 8:34 It listens to our team standup and keeps 8:37 our project tracker docs and 8:40 stakeholders up to 8:42 date. It saves a lot of time we'd 8:44 otherwise spend tracking down and 8:46 copying information across different 8:50 systems. This is exactly the kind of 8:52 thing that's suited to an agent rather 8:54 than a standard predefined workflow. 8:58 I needed to operate across multiple 9:00 contexts and make judgment calls like 9:03 interpreting something my teammate said 9:05 in standup and turning it into the 9:07 appropriate action item or 9:10 contextualizing exactly what to share 9:12 with different 9:15 stakeholders. Now this agent usually 9:18 runs programmatically. It gets triggered 9:20 from a web hook but we can also chat 9:23 with it directly. 9:25 For example, we might say, "Make sure 9:28 everything discussed today is attracted 9:30 linear and an update is sent 9:33 out." We can see the agent get to work. 9:36 First, it reads our standup transcript. 9:39 Then, it pulls in linear tickets to 9:41 understand the current project 9:43 state. At any time, we can inspect its 9:46 thoughts and see what it's doing and 9:49 why. 9:50 We can view the tools it's using and see 9:53 the exact inputs and outputs from each 9:55 one. The agent isn't a black box. You 9:58 can follow along at every 10:01 step. Here it's retrieved some data, 10:04 created a few tickets, and paused before 10:08 sending a project update. It's waiting 10:10 for my 10:11 approval. This looks good, so I'll 10:13 approve it, allowing it to send that 10:15 email and wrap up its task. 10:19 These agents are powerful. So let's have 10:21 a look at how you set one 10:23 up. On the configuration page, we define 10:26 their behavior. As a builder, this is 10:29 totally customizable. So you can adapt 10:32 agents to any scenario from simple 10:35 chatbots to systems that manage complex 10:38 processes. You can choose which models 10:40 to use. We provide OpenAI, Anthropic, 10:44 Llama, and Deepseek out of the box. or 10:48 you can connect to any other model 10:49 you've set up in 10:51 retool. We can also set model parameters 10:54 like temperature which affects the 10:56 creativity. And importantly, we can set 10:58 a maximum number of iterations to 11:01 prevent agents from getting stuck in 11:02 loops and burning a bunch of 11:05 tokens. Now, let's talk tools because 11:08 tools are actually what make agents 11:10 powerful. To do anything useful, the 11:12 agent needs secure access to your data 11:15 and systems. 11:17 Interestingly, the types of tools we 11:19 give agents are basically the same 11:21 things we've been building retool apps 11:23 on top of for 11:25 years. Agents come with lots of 11:27 pre-built tools. Google calendar, Docs, 11:31 retail storage, email, web search, code 11:34 execution, data visualization, and more. 11:38 We can easily add any of them to our 11:39 agent by selecting them here. 11:43 For the use cases that are specific to 11:44 you and your business, you'll want to 11:47 build custom tools. These can connect to 11:50 anything you've integrated with Free 11:51 Tool, giving you the ability to create 11:54 tools on top of almost any database API 11:58 or third party SAS 12:01 tool. Existing workflows and other 12:04 agents can also be used as tools. This 12:06 allows you to create multi-agent 12:08 systems. And even better, we also 12:11 connect to any remote MCP server. So any 12:14 MCP tool like GitHub or Cloudflare can 12:18 immediately be imported and used in 12:20 retool as a 12:23 tool. Looking at this agent's tools, we 12:26 have a REST query that pulls Zoom 12:28 transcripts, a set of linear tools for 12:31 getting and creating issues, and an 12:34 email tool for sending project updates. 12:38 Some tools read data, some tools write 12:40 data. For example, the create issue 12:43 tool, which writes data, expects some 12:46 inputs. We've defined those here. A 12:49 title, a description, and an assigne 12:53 ID. You want tools to be tightly scoped 12:56 with clean interfaces, and clear 12:59 descriptions. This helps the LM pick the 13:02 right tool and provide the right inputs. 13:07 This is the tool definition. The 13:09 implementation lives inside of a 13:11 function which is a lot like a retool 13:13 workflow and lets you implement custom 13:16 logic that interacts with all your 13:19 systems. Here we're making a GraphQL 13:21 mutation to create the 13:23 issue. Custom tools let you combine the 13:26 flexibility of LMS with the 13:29 trustworthiness of tested code. 13:32 The LM probably could generate this 13:34 query, but we found that getting out of 13:37 the LM as fast as possible and into our 13:40 deterministic code makes the agents much 13:43 more reliable. If we know the query 13:46 works, we want the agent to use it every 13:49 time. We let agents use most tools 13:52 autonomously. But for the send email 13:55 tool, we've required user confirmation 13:57 first. I want to double check anything 14:00 that's going to public 14:02 channels. By specifying this on the tool 14:04 level, the platform ensures that your 14:07 agents never do anything without you 14:09 first reviewing 14:11 it. So, we've seen an agent run, how 14:14 it's configured, and how it's connected 14:17 to tools. But how do we make sure it 14:20 stays reliable over 14:22 time? Agents and retool log every step 14:25 they take. You can see what happened, 14:27 what tools were used, the inputs, 14:30 outputs, and the thought process at 14:32 every 14:34 stage. This shows us the past, which is 14:37 very handy for debugging and 14:39 understanding the system. But to iterate 14:42 with confidence and track quality over 14:44 time, we use eval. 14:49 We've set up a few eval on this agent to 14:52 check that it keeps working the way we 14:54 expect, especially when we make big 14:56 changes like adding tools, switching 14:59 models, or modifying the 15:02 prompt. You can run an eval against the 15:05 data set here. I'll use our sample 15:08 inputs data set. 15:10 Each row is scored based on a reviewer, 15:13 which could be something simple like an 15:16 exact string match or more complex like 15:19 using an LM as a judge scoring 15:22 system. Once you've run a few evals, you 15:25 can compare them. After a big change, I 15:28 might compare today's version against 15:30 last week's to see where the scores 15:32 improved or regressed. 15:36 This level of observability and control 15:39 lets you confidently deploy agents 15:41 across your systems without sacrificing 15:44 trust and 15:46 reliability. As you build more and more 15:49 agents, it becomes imperative to have 15:52 insight into their behavior at scale. 15:55 With the monitoring page, you have a 15:57 full real-time view of how your agents 16:00 interact with each other, the tools they 16:02 have access to, along with other metrics 16:05 you need to make sure things are staying 16:07 on 16:08 track. That was just one example of what 16:11 you can now build with agents. 16:14 Retool Agents allows you to combine 16:16 state-of-the-art AI patterns with the 16:19 customizability of Retool, giving you a 16:22 flexible, intelligent decision-m system 16:25 on top of the trusted platform you're 16:27 already using 16:28 today. The best way to understand the 16:31 power of agents is to build one. To get 16:35 started, grab a template or build one 16:38 from scratch today. 16:41 What we've demonstrated today is the 16:43 power of building on a true application 16:45 layer. Virtual agents is the missing 16:47 piece that transforms impressive large 16:49 language models into complex automations 16:52 and real business 16:54 impact. Just as AWS became the 16:57 infrastructure layer for cloud 16:59 computing, retool is becoming the 17:01 definitive app layer for AI. And you 17:05 don't have to take my word for it. 17:07 Here's what customers have to say. At 17:09 ClickUp, we fundamentally believe in 17:12 efficiency and getting more done faster. 17:16 And this is one of the reasons that we 17:19 invested so heavily in AI from the 17:21 beginning because we saw that it was 17:24 this force multiplier that could really 17:28 really uplevel our execution 17:30 capabilities uh both internally and for 17:32 our customers. One of the first use 17:34 cases that we built out for sales was 17:37 this inbound AI agent to uh intake, 17:42 evaluate, qualify, and route and 17:45 potentially even transact uh these kind 17:48 of inbound inquiries that we have. And 17:51 over time, that's saved us, you know, 17:53 hundreds of thousands of dollars in 17:55 terms of headcount costs or and just 17:59 speed to lead. uh as well as actually 18:02 made us revenue. you know, when you 18:03 build with AI, there's, you know, 18:07 actually often 18:08 times the actual AI component of 18:11 whatever you're building is quite small, 18:13 but the business context that needs to 18:15 go into those prompts, the structure of 18:18 the outputs, the, you know, taking those 18:22 outputs and putting them somewhere 18:23 useful or or kicking off other 18:26 automations or decisioning based on 18:28 that, all of that is is what it takes to 18:32 really get value out of an AI 18:34 application. And that's the thing that 18:36 retool gives us the scaffolding and the 18:39 framework to do. And so that's where we 18:41 we really see retool as our application 18:44 layer for for AI is it really is a full 18:48 partnership across all parts of of 18:50 building an application out uh versus 18:53 just you know feeding things to AI and 18:55 and just that little microcosm. Our 18:58 vision is clear. The tools are ready. 19:00 The platform is here and the future of 19:02 work is being reimagined today. It's how 19:06 using Retool, our customers have already 19:08 automated 100 million hours of work and 19:12 how together we'll reimagine what's 19:14 possible in the AI era. Can't wait to 19:17 see what you all build.
