Parker Rex
All videos
Parker Rex

OpenAI’s Agents SDK is Here & It's Fixing My AI Startup Problems

March 12, 2025

Watch on YouTube@parkerrex

OpenAI’s Agents SDK promises a more deterministic, integrated approach to building AI-powered workflows. Parker breaks down what’s different, what’s inside the toolkit, and how it could actually solve real startup problems—not just look cool in a demo.

What’s new in OpenAI’s Agents SDK#

  • A cohesive toolkit that ties together web search, file search, vector stores, and computer-like actions into one platform.
  • A Responses API foundation that blends chat-style interactions with action-oriented tasks.
  • A dashboard and traces system to observe how agents run, what steps they take, and where handoffs happen.
  • Built-in function-tooling ideas that convert Python functions into executable tools for agents.

How agents work under the hood#

  • Tools as first-class building blocks: each file or capability (search, notes, calendar, etc.) is represented as a Tool the agent can call.
  • Looping and handoffs: an agent runs inputs, may hand off tasks to other tools, and loops until a final output is produced.
  • Traces and visibility: the traces UI reveals which steps ran, which handoffs occurred, and where decisions happened (important for debugging and trust).
  • Responses API vs. chat completion: a flexible foundation designed for action-oriented apps while keeping the familiar chat experience.

Toolkit components and capabilities#

  • Web search: internet-backed queries integrated into the agent’s workflow.
  • File search: leverage a vector store or local files to retrieve context quickly.
  • Computer use: execution of local-ish tasks or OS-level actions (within the agent’s sandbox).
  • Vector store integration: in-tool vector DBs to avoid juggling external providers.
  • Responses API: a streamlined API that combines simplicity with agent-centric capabilities.
  • Function tools: Python functions exposed to the agent, enabling dynamic tool use and orchestration.
  • Traces and dashboards: see how an agent processes a task from start to finish, including handoffs and timing.

Live demo plan and key takeaways#

  • Rebuild a Map-like, all-in-one productivity assistant using the new toolkit.
  • Create a personal stylist demo to show multi-tool orchestration: web search, file references (styling data), and a task manager.
  • Watch for the trace view: it reveals which steps were executed, which tools were used, and where a loop/handoff happened.
  • Expect real-world constraints: determinism, reliable contexts, and cost considerations will shape how you use these tools in production.

Practical takeaways for builders and startups#

  • Focus on real business value, not just flashy demos: aim for a concrete automation or decision-support workflow.
  • Don’t over-engineer: it’s easy to chase meta skills and complex architectures that don’t pay back.
  • Data and context matter: having an in-tool vector store and unified data access helps keep tools in sync.
  • Watch for security and cost: keys exposed in demos are a cautionary tale; plan for key rotation and budgeting from day one.
  • Start with a concrete use case: e.g., pull stats into a Google Sheet, automate outreach, or streamline a repeatable workflow.

Setup and getting started (high level)#

  • Install and experiment with the SDK package (exact package name may differ; refer to OpenAI’s docs for the current install command).
  • Set up your API keys securely and be mindful of local environment exposure during demos.
  • Build a small demo: define a few Tools (web search, file search, computer use) and run a simple agent loop to observe traces.

Example conceptual snippet (for illustration only):

# conceptual usage based on the video discussion
from openai_agents import Agent, Runner, Tool

web_search = Tool("web_search", func=web_search_fn)
file_search = Tool("file_search", func=file_search_fn)
computer_use = Tool("computer_use", func=computer_use_fn)

my_agent = Agent(tools=[web_search, file_search, computer_use], memory=True)

runner = Runner(agent=my_agent, input="Plan a weekly automation workflow")
trace = runner.run()
print(trace.output)

Notes:

  • The exact API and class names vary; use this to understand the flow: define tools, assemble an agent, run with an input, inspect traces.
  • The “function tools” concept is a key enabler: you can convert existing code paths into reusable agent tools.

Final thoughts#

OpenAI’s Agents SDK is shaping up to be a practical platform for building real, revenue-oriented AI workflows. The emphasis on traces, tool orchestration, and a unified toolkit could reduce the friction that typically makes agent projects feel flaky. The big question remains: can these tools drive measurable business value in day-to-day operations without over-engineering?

If you try a hands-on build, share what use case you’ll tackle first and what the traces reveal about your workflow bottlenecks.

Transcript

hey there I'm Parker Rex I led product in tech for a startup that did well over $50 million per year before we exited after that I tried my hand at launching a bunch of different startup ideas there were some really dumb ones but it coincided with all these agents coming into the scene so I was using AI every day and coding pretty much every single day last three years built a lot of different agents I'm really excited about this open AI launch uh I feel like there's been some bad ones before but this one seems different and definitely plays into the stuff that I've been building where I've struggled to kind of get deterministic answers out of agents and I just want to go through it and then also talk about this like battle between agents because there's like no code stuff that's handling certain use cases and then there's some stuff that seems Overkill to even code I don't know so let's just go through it so first of all I wanted to talk about uh the challenges that I fa is one of the products that I tried to get off the ground but ux is just crap but it's basically um and like so you further understand that I've experienced these agent problems this product is called map and it's basically like a five inone productivity thing it came out of the problem that I had where she was like I have all these goals and I wish I could just become the best version of myself if I could train a model that was the best best version of Parker God that would be really cool because it could use Biometrics to help me with my workouts and keep me on track and nudge me trained it at first to be like a David goggin's coach it was way too extreme then I had to dial it back it had didn't have any context of who I was like Beyond just like the goals that we set together um so then I started feeding it information then I was like this would be really cool if I just took every app in the dock and put it in one place so then you get the help dashboard and then you get you know your basic to-do app which is a meme in itself but I tried to make it cooler by just adding some of the stats and then having accountability on the calendar where if you had a plan for a marathon and you wanted to be really good at uh training and sticking to that then the plan would be personalized it started with health because it's more objective um but you can start to see that be a lot of information because like if you have one of these that's sending information three times a second 86444 times a day database starts filling up pretty fast um and then you know how do you get all these agents to kind of work together you need your tasks in there you need your notes in there because you know Cur for notes would be dope and I don't want to go anywhere else to then have to paste context in also had a full-blown Google Calendar clone I can't click on this full transparency because it's not running right now I can click on it but it's not going to work um but yeah it had a drag and drop basically the week view perfectly planned out and point is is like getting all those agents to kind of talk to each other really hit or miss like it clearly hasn't worked because then you know I didn't launch it um I launched a version of it but it wasn't very good and as we all kind of know like the way that these agents work and I'm hoping this SDK helps is there's a lot of stuff under the hood that's going on so as an example with like when you look at perplexity everyone's always using that and they're like tell me about Elon Musk and Tesla and so here's like a really vague query that I'm going to write but then it comes back with a bunch of links and then an answer with a summary probably something that's ranked highly on SEO because yeah and then it'll have more links and I can click on follow up and then it usually has like additional questions cool so under the hood uh this project is really helpful for kind of understanding how this stuff works but I think it's going to get destroyed um not this tool in particular but the way that we think about agents where you have all these files and they represent different Tools in here so when I go and actually make the search then it's going to call this search query it's going to have this file that represents the search and then the search might go to retrieve the retrieve might go to a query suggestion if it was an unclear query then they might do the researcher the task manager is kind of like the fork in the road so it's like as a professional web researcher um and I did the same thing with map where it was like if you didn't have enough questions or enough information about the person and they said I want to run a marathon by the end of this month like well okay well do you want help just with the running plan or do you need to get shoes or do you need a nutrition plan yada y y so all that said let's get into how they're hopefully changing it so the new toolkit that they dropped has a bunch of different stuff in it obviously um the web search it's like most of these things aren't really new but they're tying it up and making it into a nice tool kit to use which we're going to test out we're going to rebuild the thing that they had in the demo and Fiddle around with that and then I'm going to attempt to make a demo um an agent that I want which I can just quickly look at my task manager and be like here's all these things that I need to do can these uh agents do it for me so in the demo they basically speak for 20 minutes and they kind of Jump Around the Horn where web search is what it sounds like where um you know now you can give the agent your ability to reach out to the internet for queries that's cool it's not that deterministic but I guess it depends on what your output is like what's the goal of the project and um my only hope with this is that these will have real business use cases a lot of the stuff you see on the internet's like total fluff and it's like a cool research project and I've been the person doing the research project the longest so I feel like I'm qualified to say that but now I'm making money doing this sort of stuff and I just hope that people build real things with it file search this one's really interesting so it kind of ties into what they're doing behind the scenes with their dashboard their dashboard is starting to look and feel much much more like stripe where you have all these different things in here um you know it's kind of becoming this One-Stop shop for everything Ai and I really like that I hope that they don't gouge us on pricing but having a vector store in here it just begs the question of like what the heck do you need pine cone for what do you need other dedicated Vector store providers for when you just have it in here um as you can see I haven't uploaded stuff in here in a long time because PG Vector so that's the thing that would kind of prevent you from wanting to go into here or needing to go into here is because you can have it all Under One Roof postgress but this is not a postgress commercial um so file search basically the update here is now you can give an agent the ability to do that you could do that in an assistant here so this has been like the easiest most kind of rudimentary way to make one like I was learning about the Bible I'm like oh here eat this Bible and tell me answers um and you could just literally name it right give it the model add file search and then file search would reference a vector store that was done in here so the real update is just that it's now given to this toolkit which is nice hopefully it's quicker computer use pretty cool uh I'm not sure if you saw any of the demos of it when it came out but you know they partner with some companies and it's kind of like I don't know it's like a um not I can't think of the name but it's the thing that it basically takes a screenshot and then it sends it back to the vision model and it ping pongs between the vision model and the next decision and then it prints out what the action was if you're familiar with the ID extension called Klein kle does a really good job with this where like if I let's see I think I did an example of this today no they they do the best job of this if I could find an example but it's kind of like how mCP is like really hot right now and they're like oh my gosh I can get logs and you're like yo you could have done that before with Klein but they'll take a screenshot of what's going on and then ping pong it back and forth between the vision model and answer that and I think that's really awesome because now it's in this tool so I think what they're doing will give devs this full toolkit where you can tie all them together the responses API they basically just beefed up a new API but let's let's look into it I'm not the master I'm just trying to give you the summary um where's the dock responses API I know we're going to want that one and then I know we're going to [Music] want responses I think they'd put a little star next to it or something it's probably on their blog what's new responses API okay the responses API are two ways to interact with open eyes models great so the responses API is not full feature dependent or feature matching but combines to Simplicity of the chat completion with the ability to do more agentic tasks as the model evolves the responses apis a flexible foundation for action oriented action uh applications with Y those tools built in it's not going away don't get scared compare the code okay so left side right side what I did notice is just that the the vernacular like the the syntax is a lot cleaner when I was looking at examples of it but like for instance this one web search so you just basically put in a weight a variable then you pass it what the tools are that you want and then input that you want and then you log it simple enough file search you are able to create the vector store that's nice I don't know I've never done that VI a code full transparency but it makes sense because in map world we were spinning up a new IND for each user so this is the same thing okay um you can pass it the tools Okay computer use browser I wonder what the variables are for environment I saw Linux earlier so I was kind of confused because why would I guess they have spinning up a VM and then you could say oh I want yuntu not sure there no way the differences so okay key improvements make it flexible okay cool I'm going to start building something I'm getting getting over just looking so let's grab our little pen here because I want to scribble some stuff so we obviously did all that stuff we did that we did that we did that we did that we want to look let's look at this real quick um that's DK enables you to build generic agentic AI apps and a lightweight easy to use package so yeah basically they just took swarm and then they beat it up they allow for looping python first handoffs got it okay this was what was really cool was the tooling behind the scenes basically lets you see what happened in a guey that's awesome and the function tools is another big one that they mentioned where basically like they just read the python function and turn it into whatever the thing needs which is really cool so we're going to install this and run it in a sec um what was the other thing I found interesting on there oh this is just commentary where I feel like a lot of people that are building right now um and i' again I've been that person and I just want to call it out that they'll build stuff then like become the expert in some like meta skill that is valuable for all of like a month and a half and they'll build all these crazy systems and I don't want to name names but especially on YouTube where it's like you're over engineering stuff like and it's not making money it's not saving you time so don't do that so let's get those ha look at super base complaining you're so funny let's make a little directory make their oi you new SDK oi new SDK pip install open AI agents there's a guy who really just uses typescript all the time um you know I've I've gone out on the edge and tried to write other stuff but I always get freaked out when I used pip because I've messed it up so bad in the past where then I'm like I gota use pip X or whatever but seems like it it worked oh no AER is mad but that doesn't matter cares so let's do uh file and let's open up and Sir d r sir gosh so that that's funny I downloaded a I updated my Zs rhc the YouTubers said that this new thing would be faster but it didn't even open so he was wrong let's look we're going to go through the same demo that they did they built a stylist a personal stylist and they uploaded all these files of like what people wear at their company and it was I guess a good use case because it covered all the bases of all the different tools but it's something no one would ever do which I think is weird but basically The Stylist was able to figure out what the person would want to wear and then you know look up shops where they could buy a Patagonia vest in a city so interesting choice now we want to hand type this because we're crazy from Agents import agent and Runner okay easy enough what else is you can I wonder what else there is Agent hooks oh is that the computer computer okay computer tool I just want to know just want to know what's in here run a workflow starting at a given agent the agent will run in a loop until a final output is generated the loop looks runs like so the agent is invoked with the given input if there's a final output it terminates if there's a handoff we run the loop again so handoffs are what you would see in this new fancy uh oh wait a second I thought it was in here no this is the dashboard you'd see it in this new traces thing so they showed a demo of it after the style the personal stylus ran and it shows you basically like which steps got handed off to which did not and it goes back to the combo at the beginning when I went through morphic and I went through map where you know you have specialized agents for very specialized tasks um then you you know chain them together Chain's even a bad word now so let's get that demo back up I got to get an API key one second let me move this off obviously I need that yeah and you're just getting a hello world great that was awesome it worked so let's see what the trace was cool we got a trace no handoffs and it took two seconds what a good job then we can click in here you can see that it hit the assistant I can also see that I totally uh showed my environment variable regardless so now I gota change that that's awesome let me remind myself of that even though probably three people are going to watch this but change or rotate API Key View date and one hour well I kill for die Coke right now so I'm scratching my neck okay so I can see the token count I can see the response I can see all the stuff see it's not that interesting when it's one thing but it'll be more interesting when there's multiple things so let's make multiple things uh yeah I said poor example I said that I said that and I said that and I said that good so now let's think of a good idea uh I want to make an agent that can go to I don't want to do anything with off I don't have to how about I want to use computer use how can we combine all of these things together let's just think of it hey I'm a content creator and software engineer and I want to test out open ai's new agent SDK that includes the following tools web search File search computer use responses API and agents SDK and there's a couple things that I know I need I need um the ability to do the ability to pull all my stats for all my different brand pages into a Google sheet that's one use use case potentially another I'm just going to Ramble On use cases and then come back with something save me here um I need to do automated Outreach for potential clients I need to do um I want to be able to turn my eight sleep up or down or off that would be a poor use oh man fellas because there's no girls watching this um this is tough fellas I'm just going to ramble let's that's it [Applause] fellas this is not a video of me building one because I can't think of one I don't want to sit here and come up with something really bad sometimes sometimes think of something that can come up with all five thanks for watching I'll see you in the next one oh man