OpenAI’s Agentkit Just Made Building AI Agents EASY — Here’s How!

OpenAI just launched agent kit, a complete set of tools to build, deploy, and optimize AI agents. If you ever stitch together orchestration, connectors, eval pipelines, prompt tuning, and then spend weeks on front end before launch, this release is meant to compress all of that into something faster and simpler. In this video, I'll walk through every section of OpenAI's announcement, agent builder, chat kit, and the new eval capabilities, reinforcement, fine-tuning, and pricing and availability, all straight from the official release. So, let's get into it. So, why did OpenAI decide to release the agent kit? Well, number one, right now, OpenAI believes that when you're trying to make agents, you have to juggle fragmented tools. You have complex orchestration with no versioning. You have custom connectors. You have manual evaluation processes, prompt tuning, and weeks on weeks of front-end work. Why waste all that time when you can use their agent kit, which addresses this with three building blocks, so you can design your workflows visually and embedded agentic UIs much faster. Number one, they're providing you with an agent builder, which is kind of like their visual canvas for creating and versioning multiple agent workflows. as well. There's a connector registry which is going to be your central place for admins to manage how the data and tools connect across the AI products you might use. And lastly, there's a chat kit which is a toolkit for embedding customizable chatbased agent experiences in your product. If you're enterprise with thousands and thousands of customers, you can reduce your cost by just using OpenAI's agent kit as well. using jacket which is something that most enterprises are looking for anyways can help them save time as well cut cost. So this is something big as well. Open AI is also expanding with evaluation methods especially with data sets trace grading automated prompt optimization and support for third-party models. All of this is to measure and improve how an agent performs. So they're not only providing you with a toolkit that helps you build it but also providing you with a evaluation method. and we'll get into that later on in the video. Let's take a look at their agent builder, which is kind of like a visual canvas that allows you to use logic and drag and drop nodes, connect tools, and configure custom guard rails. And you get to kind of visualize all of that process. And it makes the whole process of building an agent much easier because it's visual. You also get preview runs, inline eval configuration, and full versioning, which is ideal for rapid iteration. Let's take a look at an example of how this is used. In this example above here on screen, you can see how the agent canvas works. You have the ability to start off with a basic simple layout and then you have a classifier agent that you're allowed to obviously give it a name, instructions about what it is, as well. You can kind of choose what kind of contacts it's supposed to have. Full conversation or not, what model you're using to run that agent, and the amount of effort your agent is supposed to put in to get you an output as well. You can connect more tools, determine what the output format should be. Should it be text, should it be more visual? We can see that like this is how you start off and you can go from here. Once you kind of get a hang of it, you can add a little bit of logic. So you still kind of need to understand coding logic, need to understand how you would think about a problem logically and kind of use that agent canvas to build that logic. And then you can see that once you start building that logic, you can break it off into like okay if the input is XYZ, let's go to this step or this step. You also get to see how the model is thinking, how the agent is thinking and solving the process. Open AAI also worked with real companies and got some feedback. Ramp said that agent builder transformed what once took months of complex orchestration into custom code and manual optimization into just a couple of hours. So we can see this process is making things much faster. The visual canvas keeps product legal and engineering on the same page slashing iteration cycles by 70% and getting an agent live in two sprints rather than two quarters. So these are some huge results and enterprises really value their time as well transparency. So by using something like the visual canvas it keeps different departments on the same page which is a huge win. Another example is ly corporation which said agent builder allowed us to orchestrate agents in a whole new way with engineers and subject matter experts collaborating all in one interface. So the visual canvas really makes it easy for cross department to collaborate on one page which is another big win for enterprises. We built our first multi-agentic workflow and ran it in less than 2 hours dramatically accelerating the time to create and deploy agents. So OpenAI has clearly made making agents much more simpler, much more transparent, much more collaborative and most importantly much

more faster. Another key concern for enterprises is governance and safety when using these agents. So, OpenAI has provided enterprises a connector registry, which is a central admin panel that consolidates data sources across chat GPT and the API, including pre-built connectors like Dropbox, Google Drive, SharePoint, and Microsoft Teams, plus thirdparty MCPs. And you can enable guard rails in agent builder like the ones we see here and here. The guard rails can mask or flag any jailbreaks and apply other safeguards. The guardrails can also run standalone or via libraries for Python and JavaScript. Another key problem for enterprises is deploying chat UIs for agents is sneakily hard. You need streaming responses, thread management, show the model thinking, and in chat experiences. Chatkit makes it straightforward to embed chatbased agents into apps or websites and customize the user interface to match the brand. Let's look at some examples of how enterprises have benefit from that. Number one, we have Canva. Canva claims that they saved over two weeks of time building and support agent for our Canva developers and integrated in less than an hour. Supposedly the support agent will transform the way developers engage with their docs by turning it into a conversational experience. Let's look at another example from legalon which says that by adopting chatkit we develop UI for an AI agent in a day reducing development time cost by as much as 80%. Which is huge. Hubspot said made it easy for us to prototype HubSpot's breeze assistant. We saved weeks of custom front-end work and deployed an agent that dramatically improves time to resolution while giving higher quality support to our customers. So, we can see that enterprises are not only benefiting from the agent kit, but they're also able to use chat kit and deploy it for their businesses very quickly. Another thing to keep in mind is to ship reliable agents, you kind of need to make sure that they're evaluated properly. And open a previously launched some evaluation methods, but now it adds three new ones specifically for the agent kit. Number one, data sets. You can build agent eval from scratch and expand them over time with automated graders and human annotations. For this example, we can see that it provides some feedback as well as some accuracy scores. Number two, we see prompt optimizer, which are automated prompt optimization that allows you to improve your prompt from a human grader perspective using a plus or minus signal. As we can see up here, plus two, minus two depending on the prompt you give it as well. There's trace grading. Trace grading assesses full agentic workflows and autograde to pinpoint issues. So we can see that they assigns a grade to it and as well it gives a fail or pass depending on the evaluation method use. OpenAI also says that customers are already seeing gains from this. For example, Carlilele cut development time on multi- aent due diligence frameworks by over 50% and increase agent accuracy by 30%. Which means higher speed, higher accuracy and that equals to less iteration cycles. So that's the agent kit in a nutshell. A full toolkit to design, evaluate, and launch your own AI agents all within OpenAI's ecosystem. It's another major step towards making Agentic AI actually usable at scale at enterprise level. If you found this breakdown helpful, hit like, drop a comment on which part of aging kit you're most excited to try, and don't forget to subscribe. Until then, see you next time.