# AI Agents for Economic Research **Channel:** VoxDev **Source:** https://www.youtube.com/watch?v=YPv9BqweQIo **Transcript page:** https://www.withtranscript.ai/video/YPv9BqweQIo ## Chapters - 0:23 — Webinar Introduction and Talk Overview - 1:33 — Why Economists Should Care - 7:06 — Choosing Between Codex and Claude - 11:28 — First Exercise: Clean Up Old Code - 19:40 — Avoiding Beginner Mistakes with Git - 24:26 — Using Skills to Guide Agents - 32:58 — Planning Before Implementing Work - 35:43 — Learning and Budgeting for AI Tools - 41:25 — Q&A: Privacy and Sensitive Data - 44:52 — Q&A: Security and Sandboxed Use - 47:36 — Q&A: Collaboration and Dropbox Workflows - 49:47 — Q&A: Tool Comparisons and Cursor - 52:36 — Q&A: Hardware and Memory Needs - 54:51 — Q&A: Finding and Customizing Skills - 58:06 — Q&A: Copilot for Students - 1:00:59 — Closing Remarks and Farewell ## Transcript **[0:23] Speaker A:** Okay, great. I think we'll get started there. Thanks everyone for joining today's webinar um on getting started using AI agents for economic research. I'm Oliver Hani, the managing editor here at Vox Dev. Um we are really excited to be joined by Aniket Panjwani, who is an economist himself, who did a PhD at Northwestern, and is now a director of AI and ML at a fintech startup. And he's been helping other economists learn how to use agents for their work. If you have any questions throughout today's presentation, we aim to leave a lot of time for questions at the end. So, please use the Q&A function on Zoom, and I will pose as many as possible to Aniket in the time we have. Aniket, uh the floor is yours. Awesome. Let me share my screen. Uh there we go. **[1:17] Speaker A:** Okay. Are we good, Oliver? Yeah, all good. Thanks. Great. Um great to be here. I'm going to be talking about AI agents for economics research, some practical workflows, budget trade-offs, and why economists should care. **[1:33] Speaker A:** Um did my PhD at Northwestern, as Oliver mentioned, and currently work doing AI and ML work at a at a fintech startup, Payslice. So, I wanted to start with this tweet from a fellow named Dylan Patel, who's uh fairly big in the AI space. If you're an economist, you may not have heard of him. If you're uh in AI and agentic coding and following what's going on right now, you've definitely heard of him. So, he says here, uh "Being in SF is like being in Wuhan right before the pandemic. Something is happening. It's going to hit everywhere, but so people few people know it." And so, okay. I mean, that's uh you somewhat know what he's talking about, but maybe it's a little bit cryptic. But, what is it exactly perhaps that he's talking about? Well, 2 days later, so this was just 4 days ago, 2 days later he gives a pretty big podcast where he says, "First, you already see the junior software engineering market is nuked. **[2:34] Speaker A:** And so, this is something that, you know, most people are aware of, but maybe that's not that relevant to economists. And he's also saying, you know, you look at Claude code spend, right? It's it's freaking nuts. And this is something that you're seeing in business, the amount that people are spending on these tools is quite high. But, he also says, "Claude code is not a coding thing. It's for all people." And so, maybe my initial question here is, are economists included in all people? And my answer is a definitive yes. So, let me show you some examples of this. So, this is Paul Novosad, um a development economist at Dartmouth, who says, "I wrote a decent paper with AI. It took me about 3 hours from start to finish, including an interactive choose-your-own-border RD regression discontinuity adventure. It's a what are we even doing here kind of day. **[3:38] Speaker A:** And this discussion, all of these links will be available. I'll put them on my site, and then they'll also be on the Vox Dev site. But, the discussion here is interesting to follow because in his opinion, and I didn't read the paper in detail, but in mine, too, it's already the kind of research that would, maybe with some cleaning up, be satisfactory at like a decent field journal 5 years ago. But, you know, all sorts of expectations are going to change with the introduction of these tools. And then, here's Gauti Eggertsson at Brown saying, "AI is a game changer for economic research. We will look back and think before and after." The junior job market used to place enormous value on technical skills, but now you can just experiment in days with tools like Claude Code, or just in an evening. And so, that's another way in which economics research is changing with these tools. **[4:37] Speaker A:** The cost of experimentation is greatly decreasing. And here is this Chris Blattman at the Harris School of Public Policy at the University of Chicago. And the reason I brought this up is to highlight how quick your personal workflow transformations can be once you start using these tools. So, 4 weeks ago, he's a Claude Code skeptic, not a coder, none of the use cases are relevant, but he's created now a set of tools to help him with all the types of big project management tasks that somebody managing field experiments and students and big teams has to deal with. And so, these are a few of the ones that he's created. A tool for continuous improvement, a tool for handling his taxes, a data analysis co-pilot and how to teach AI your voice. So, how to use it effectively for writing. And on his website, Claude Blattman, he's making all of these publicly available for anyone. **[5:39] Speaker A:** And plenty of economists that I've trained at various universities across variety of fields, finance, macro, political economy, Alexy, and Hayeus more economic history have gotten great benefit from Claude Code after just short sessions working with me. And one last piece here, which I want to highlight, is that is this is Claude Code and agentic coding in AI just helping us with empirical research or will it help us with theoretical research? And so here I'm bringing up a mathematician Terence Tao, who is one of the preeminent mathematicians in the world today, a Fields Medal winner. And what he noticed is that this that an older version of GPT, GPT 5.2 Pro, was able to more or less autonomously by AI solve an Erdos problem. So a class of problems, open problems, which had been posed by the mathematician Paul Erdos. And so to the extent that there are analogs to what mathematicians do to large degrees of theoretical research, it's pretty clear that even if you're a theorist who's not working with data, who's not coding, it's incumbent on you **[7:02] Speaker A:** to become familiar with and maximize your use of these tools. **[7:06] Speaker A:** And so in this talk, I'm going to keep the scope pretty tight. First, I'm going to talk about what agentic coding tool you should use, and I'll tell you what these are in the first place. I'm going to give you a practical first exercise that immediately after this talk, you can go and do right away today, and which will be very helpful to you. I'm going to talk about three beginner mistakes that I see economists or just agentic coding beginners in general making and how to avoid them. And I'm going to give some suggestions on how you can keep learning agentic coding for economics research purposes. So, let's start here. What agentic coding tool should you use? So, there's really just two options in the right now. There's Claude Code or Codex. Happy to answer questions about other options, but for practical purposes, these are the only two that I'd consider. **[8:03] Speaker A:** And so, I ran a poll on Twitter a week ago and I asked economists what agentic coding tool are they using. Got 439 votes and most people voted Claude Code. And but my suggestion is to at this point, especially for this audience, to not use Claude Code. So, my understanding is many of the people watching may be new to this space, perhaps skeptical, perhaps not, but you may have a tight budget. And so, let's say you have $20 per month. With that $20 per month, which tool should you use? I recommend Codex for a few reasons. First, the GPT-5.4 model right now, which is what you can use through Codex, is an order of magnitude better at pretty much every task other than web design than Opus 4.6. Second, quite important, even before the recent generations of Codex, like for the last 6 months, it's simply been the case that Codex at $20 per month gets you a lot more usage before you hit your limits than did **[9:14] Speaker A:** Claude Code at $20 per month. And so, that's going to be pretty important for you guys if you're on a $20 a month plan. Third, the Codex desktop app, which is a desktop application which accompanies uh the Codex coding agent, is currently the best agentic coding interface for 98% of users or more. And so, that's something that I would recommend you use. Now, I've been using Codex for 95% of my work for the last month. However, what you saw in the previous survey is that it seems like a lot of the economics profession is converging on Claude Code, and even I was using Claude Code daily, you know, between something like February of 2025 when it first came out until August of '25, then switched to Codex, then back to Claude Code, now back to Codex. Will it be difficult for you to work with Claude Code users if you're on Codex? **[10:22] Speaker A:** And the answer is no. So, one thing which I'll be talking about later in this talk is a concept called skills. And so, skills are a sort of way to give these agentic coding tools specialized capabilities. And it's very easy to transfer from a set of Claude Code skills, which most economists are using, to Codex skills. So, that's a solved problem. Also, it's very easy to change sort of the base set of instructions that you're giving on a project-level basis, which is what this claude.md file typically has, to Codex. However, as I said, this space changes quickly. I've switched quite a lot, and so, if you want to use Claude Code, if you're already using it, the rest of this talk is equally applicable regardless of which of these two tools you choose to use. So, that's the initial decision, especially for those of you who are on a tight budget and who have not started using these tools yet, I recommend you go with Codex. **[11:28] Speaker A:** So, now I want to give you a practical first exercise that you will be able to start with today and which will be very useful to you. And that is to clean up your old code. So, uh if you do any bit of self-reflection, if you're an economist here, I would say 98% of you are also going to realize or know that you are horrible programmers. And just to emphasize that point, I have another bullet point where I just want to say how horrible horrible horrible of programmers you are. And in my opinion, one of the greatest boons that this agentic coding revolution is going to give us is that we'll never have to see an economist-written line of code ever again in our lives. And so, one great way to get a good benefit out of these agentic coding tools on day one very quickly is to let coding agents reorganize and rewrite your project directories for you. **[12:33] Speaker A:** So, that both includes reorganization and rewriting your code. But, how do you do this in a non-destructive fashion because you don't want to like lose all your work by accident? And how do you validate that what the coding agent produces for you is correct? So, first, I'm going to give you some prerequisites of things you need to do step by step in order to get to the point where you can do this exercise. And then I'll tell you how we're going to go about meeting these conditions. So, the prerequisites is first, choose your tool. I've suggested Codex. If you got Claude Code, you can use that. Next, you have to install Git, which is the version control tool which anybody using version control uses. Most economists don't, but as I'll talk about later in the talk, it's very important and useful to be using Git alongside these agentic coding tools. **[13:30] Speaker A:** Then I want you to create a GitHub account, which is distinct from Git, but a very useful tool to host your repositories remotely. And I want you to install a tool GH, which is a command line tool, which then your agentic coding tool will be able to use to interact with your repositories on GitHub. And then finally, I want you to set up Stata since I'm imagining a lot of you are using Stata with either Claude Coder or Codex, and I've got an article here with the corresponding video in which I've explained how to get Claude Code set up to work with Stata. Okay. So, what once you've done those steps, I just want you to go in a terminal, navigate to your project folder, or if you're using an editor like VS Code, open up Claude Code just in that project folder, and then you can just take this You can take this entire prompt here, copy it, paste it into Claude Code, and it should fairly seamlessly for even fairly complex projects, even though this is a pretty naive approach, do a **[14:49] Speaker A:** good job of helping you clean up, refactor, and like rewrite a lot of your code in a safe way that meets those two requirements of, you know, being safe and validating your work. But, I want you guys to understand things as well, so I'm going to walk through a little bit each of these lines and why I'm including them. Because a lot of getting good use out of these tools is being able to prompt them effectively. So, the first line here, I'm saying, I want your help cleaning up this project and setting up a Git repo." So, that's the former version control we're using. And then, copy it all over to a backup directory. So, okay. Even if this like completely messes up everything in my codebase, at least I have a backup. The next step is to suggest an initial Git ignore and create a first repo. **[15:45] Speaker A:** So, what is a Git ignore? That's a set of files in your Git repository that you decide not to commit to your version control system. And so, the typical types of files that should be included in there as an economist are your data, if it's really big, which often it is, and the outputs of LaTeX, I imagine you're using, like the PDF. You want to commit only the actual tech document and not the PDF. Cloud Code is smart, and so it's going to know these things and many other things that shouldn't and should not be committed, and it will help you do that. Next, push the initial repo to a private repo on GitHub. So, you probably don't want your entire project to be visible to everybody in the world. So, if you just specify private repo, those are free to have on GitHub, and Cloud Code or Codex will help you do that. **[16:46] Speaker A:** And finally, here I'm asking to check out a new branch, so that's sort of an independent history of your project on which we can now write, but also either easily combine back into the main history, or just keep working on independently, decide we don't like it, and revert back to the original history. And suggest how to restructure the repo to have best practices with data, code, docs, paper organization, and lineage. If you have other things that you want it to do, you can pack it into this sentence. Now, one option is that you could just copy this and paste it, and then what Claude Code will do is it will since we're asking for a suggestion, Claude Coder Codex will come back more with a plan of what to do, which you can then approve or deny. But, if you're confident with these tools already and their abilities to make choices, if you paste this whole thing, it might pause here for you to give your input, or it may just go through and do the whole **[17:53] Speaker A:** thing. I'm not sure. I haven't tried this recently, so you know, you have to test it out. Then, just do it on the new branch. So, here I'm suggesting like figure out how to restructure this, and then do that on the new branch. And then in the last sentence here, I'm saying, "Finally, actually run any code you've modified or restructured to make sure it lines up with the prior results." And so, when economists have messy directories, I found that Claude Code is pretty good at finding the versions of the code which are like antique, and the versions of the code which are current. Because, from what I've seen, people will just have, you know, uh _v2.do or _v7_f- um excuse me, final.do. And so, Claude Code is pretty good. Codex is pretty good at seeing through these bad programming practices, finding the real version that needs to be tested. **[18:54] Speaker A:** And it'll ask you first, "Is this the right one?" But then, it can, if you've set up Stata correctly, if you're using Stata, actually also run your code, look at the output, and then tell you if it matches between your source directory, your original one, and this refactored one. And so, that's a a nice first step that, you know, if you wanted to cut this talk off right now, you could just go install Codex, do those prep steps I mentioned, copy and paste this prompt into some of your directories in the Codex window, and you will get at least $20 of value out of this tool. Okay. So, this is a nice exercise that, you know, I would recommend to you to start. **[19:40] Speaker A:** But now, I want to discuss a few different mistakes that I see beginners making, and how I would suggest avoiding them. So, the first mistake is to not use Git, or to treat it as optional. And this is problematic in a few different ways. So, Git, in my opinion, makes these tools 10 times as powerful. But you don't need to know Git to use Git. You just have to tell the tool to use Git, and then it's going to know, pretty much baked into its training, best practices with Git, how to form commit messages. It does help to have a vague notional understanding of Git at the very minimum, in order to be able to more effectively communicate with the tools, but uh you don't have to know it at a deep level to get productive use out of it. And so, one thing I recommend is perhaps to just put in your sort of agentic coding global instructions, a simple line like, "Make atomic Git commits as you go along." **[20:59] Speaker A:** And so, this is an instruction that will be invoked anytime you start a Claude code or Codex section in a particular project and that's going to help you, you know, get a lot of value out of Git. And so like what what does Git give you? Why is it so important? So, one thing it gives you is the ability to safely try new things and roll them back. And so, one of the great things about these tools is that they can work autonomously often for half hour, hour the types of things that economists are doing from what I've seen working with economists like on data cleaning or data analysis type tasks, you're not going to get autonomous use of these tools for that long. Those may stay more in the 5 to 10 minute range, but for people who are doing structural estimation if you sort of give the right plan to the coding agent I've had it go off and just work for **[22:03] Speaker A:** half an hour, 45 minutes, an hour and come back with very good results. But if you don't like what you get, you want to be able to roll it back. And so that's an important thing Git gives you. Another thing Git gives you is that it gives your agentic coding tool memory because these tools know very well how to use Git and so part of what they're able to do is to look through the history of changes that have been made on a project to help you understand, for example where a particular error may have been introduced in your code that's leading to some weird result or weird figure that you can't understand. And so that's another important reason to use Git. And a final sort of more advanced usage of these tools, but something which the codex desktop app makes very easy for you to do is that it allows you to use a concept called work trees. **[23:01] Speaker A:** And work trees allow you to work on multiple aspects of a project at the same time. And so, the problem is that once you get more advanced with these tools, you'll often want to have multiple Cloud Coder, multiple CodeX windows open at the same time. But, if they're all working on the same set of files, that's going to be problematic because they'll overwrite each other or there'll be a change made and they don't know where it came from. So, that can confuse them. And so, what you can do instead is use this sort of get convenient get command to create an entirely different copy of the project in a simple way, but like have an understanding of where each of these copies is relative to each other so that the tool can help you combine them together in a seamless way when the time comes for you to do that. **[23:53] Speaker A:** And so, those are three quite important things that Git gives you if you want to understand and get better. So, I think something I forgot to mention is I make a lot of YouTube videos on Cloud Coder and CodeX, some directed at economists, some not specific. But, uh if you want a better understand Git, I've got a uh a good video. That's like a complete beginner's guide on this. And all of these are linked, including all the Twitter posts or whatever I show you, so you'll be able to go directly and uh evaluate my my video for yourself. **[24:26] Speaker A:** Okay. A second mistake that I see beginners making is that they underutilize skills. And so, skills are something that I mentioned a little bit earlier, but let me give you the definition of a skill in simple language again. So, skills are essentially reusable instructions that teach CodeX or Cloud Coder how to do a task in your way. So, a way to think of them is that they're saved playbooks. So, like whenever you do something, like you want to try triage your email inbox, or anytime you want to write a particular grant for the NIH, or um anytime that you want to uh create um homeworks from your lecture notes. You have a set of ways that you want to do it, and you can encode that you can codify these in these skills. And so, anytime that And so, the agent is able to start with these default set of instructions every single time. **[25:30] Speaker A:** So, let me show you some real skills that economists have been making, and you can study them and take a look at them, and I'll show you how you can, immediately after this talk, within 30 seconds or a minute, start making use of these skills as well. So, here's a paper from Pedro Santana, who's an econometrician at Emory, and he's made a skill called review paper. And so, one thing you'll see, this is the beginning of the skill, is every skill starts with what's called front matter. And so, the front matter is a set of three or four fields that must be in a skill for the agentic coding tool to register it and start using it. And actually, I think only two are needed, the name and the description. So, the name is review paper, and the description is comprehensive manuscript review covering argument structure, econometric specification, citation completeness, and potential referee objections. **[26:31] Speaker A:** And so, anytime that you load Claude Code, um Claude Code will automatically have this front matter in its context. But then, a skill also has a body. So, there's like other text that gives the full set of instructions in the skill. So, here are like some of the lines that are in the full set of instructions in Pedro's skill, but those full set of instructions are only invoked if the agentic coding tool thinks that this skill is relevant at the point in time given like whatever prompt you've given it. So, if you tell it, "Hey, I got a referee report to do. Can you review this paper?" then it's going to read the rest of the instructions. And the reason that's important is because the problem with these tools, agentic coding tools, is that they have a limited set of tokens or words that you can make use of. **[27:27] Speaker A:** And so, that's the reason why skills are important. It's a way to sort of give the first set of instructions that um or like an indication of a useful set of instructions to the tool without it needing to load in the entire set of instructions every single time, many of which may not be relevant to whatever task it is you want it to do. And so, this is a review paper skill, which if you wanted to use, you could just click this link, copy that URL, and then paste it into your agentic coding tool of choice, whether Claude code or Codex, and just ask it, "Hey, install this skill for me." There. Now you have a great uh review paper skill. Another one which Pedro has made, which I imagine will also be useful for many of you, is a data analysis skill. So, Pedro likes to use R. **[28:20] Speaker A:** I would recommend moving to R or Python rather than staying with Stata if you don't have strong switching costs, but many of you do, so you'd be able to recreate this one in Stata if you wanted. So, what's the description? It's an end-to-end R data analysis workflow from exploration through regression to publication-ready tables and figures. And then these are some of the lines that are exactly in here. And so, the way he has it set up is that the body of the skill is split into phases. One for setup and data loading, two for exploratory data analysis, three for main analysis, four for publication-ready output, five for save and review. And so, a lot of like times you would it would be very useful to have just a quick and dirty analysis, which ordinarily you would have some RA go off and spend 2 weeks on. **[29:14] Speaker A:** Now, like it may take in the amount of time to do like a quick and dirty analysis an hour, hour and a half for the tool to run. But, from your perspective, all you have to do is invoke the skill data analysis, point it to the right files, and then, if you've set up your environment so that CodeX or Cloud Code can work autonomously without asking you for instructions every time, just go get a coffee, go play with your kids, or do more work, do different work, do other work you have to do. And so, that's a sense in which these tools let you sort of parallelize your work, but also, you know, aspects of your life. And now, the last one I wanted to show is one from Chris Blattman called proposal write, which I imagine is like quite useful to a development economics audience, right? **[30:07] Speaker A:** And one thing I wanted to point out about Chris's skill, this is not to put him on blast, is that this skill here is not a proper skill. And so, a naive mistake that people will make is that I was working with an economist, for example, who said, "I made all these skills, but Cloud Code can't find them." And so, in order for a skill to be recognized, it has to be put in a particular location on either your computer, if you want it to be visible to all repositories anytime you're working with it, it needs to be or in your particular repository if it's scoped to just that project and it needs to have the front matter. So it needs to have a name and a description. So this one doesn't have a name and a description. So even if you put it in the right place, Claude Code would not find it automatically. **[30:59] Speaker A:** But it's an easy thing to fix and you don't have to fix it yourself. You can ask CodeX or Claude Code to just do it for you. But the structure of the workflow is also like very interesting to read. So these are some exact lines. The skill produces a first draft of a funding proposal in markdown. Here's the flow he has in mind, an input checklist, auto gather context, donor profile, context brief, analyze and plan, draft and save. Present this checklist immediately and collect all inputs before proceeding. So that's a very good and useful technique that if you want to find ways in your skills as you iterate on them to find guardrails to be able to check its work before it goes on and does later work. And so that's another useful skill. Now to make your own skills in CodeX there's a built-in skill creator skill that you can use. **[31:53] Speaker A:** In Claude Code you have to install the skill creator skill. It doesn't come by default at least the last time what it checked. And then you just tell it what you want to make. So you just tell it the skill you want to make and it'll help you. Like you don't have to think too much about it. And especially for skills that don't interact with external systems like email skills would. Um but which are very useful, they're very easy to make and Claude Code can make you something useful, CodeX can make you something useful and you know the in in a few seconds. And so here is the Claude Code skill if you want to install it um and you can just ask Claude Code, "Hey, install this skill for me." and it should be able to do do Okay. And I've got a video here which I cover in particular this skill creator skill in detail. **[32:44] Speaker A:** I like walk you through every single piece of it, and I would recommend watching this video because it will help you get a good conceptual understanding of what skills are so that it's less of a black box to you. **[32:58] Speaker A:** The last mistake that I see uh beginners making is that they don't make plans before they do their work. And so, a problem with coding agents is that you have this context window in which you have to work. And uh often times when you're making a plan, the things that go into making a plan, like the ideas you had and the brainstorming you had and the false lines you went through while thinking through a plan, are not actually relevant for the implementation of a plan. And as that sort of like thread gets bigger and bigger, there's more irrelevant information to the final implementation of a plan. And so, these long threads get noisy. And then, as you get towards the limits of what you can have in one coding agent session, you'll notice the performance dropping. So, like people will say uh Codex got dumber or Claude Code got dumber. **[33:53] Speaker A:** That's what's happening. So, if you write a plan first for complex things you want to do, that's going to protect the context window because what it allows you to do is take that plan, write it out to a file, and then start a new fresh session where all you tell it in the next fresh session is just implement the plan which I've written out here. And so, each of um Codex and Claude Code have what is called plan modes which facilitate doing this. And there are also many plugins, like the Compound Engineering plugin is popular with a segment of economists which um have their own sophisticated versions of like optimizing the creation of plans. And so just to the sort of give you an understanding, like the naive approach is to do everything here in one long thread, but then, you know, all the stuff that goes into planning may not have been important **[34:55] Speaker A:** for implementing. Whereas alternatively, you have a single thread in which you plan and you write out to a markdown file plan.md, and then you have a fresh session in which you execute that plan. Or if your plan is done partially and you still need to refine it, you can also, you know, go there and then start a new session and then uh iterate on the plan. And so this is like a very simplified version of what I'm describing. There's a lot of like ways that you can optimize on what I'm describing to you right now to uh get the best results out of these tools, but it's an idea that I wanted to be in the forefront of your heads as as beginners as a common mistake that beginners make. **[35:43] Speaker A:** Okay. So, how do you keep learning agentic coding for research past this talk? I'm going to give a few pieces of advice. So, my first advice is to get on X. So, there is some serious discussion on agentic coding for economics on LinkedIn. There are like some economists who will cross-post between LinkedIn and X, but if you wanted to compare early adopters to late adopters, the early adopters are just the ones who were on Twitter on X. Uh Blue Sky is a total dead end. Like there's nothing going on there. There's a lot more discussion among economists on X, and all the discussion among developers is on X. And the reason why that's important is that everything that economists are learning is just downstream of like what is happening among developers. And so if you want to be on the cutting edge, like you want to sort of position yourself somewhere in this early adopter space. **[36:40] Speaker A:** You don't have to be the earliest adopter, but it's nice to know like when something is sort of hitting particular early economist adopters so that you can adopt this yourself. And so I made a post here which I wanted to link on everyone I follow to keep up with on AI. It's kind of small here, but it was a popular post. Got 150,000 views. But then I also made a a list here which you'll be able to get in the slides of economists who I think are putting out good stuff specifically on agentic coding. On AI it's much broader, but like each of these has is either doing some cool experiments, talking about workflows, or something like that. So these are a set of people that I would recommend you follow to keep up with this space. Now, last topic I wanted to talk about is the money question. **[37:29] Speaker A:** So like I know that a lot of people here have tight budgets, $20. And so like how is that going to limit you compared to somebody who has a bigger budget? Well, everything I said in this talk like can be done on that $20 budget. It's just that with that $20 budget, you're going to have limits both in a particular day, in a particular 5-hour session, you have a certain limit of how much usage you can use it, and on a week. And you're going to hit your limits a lot faster if you're on $20 compared to if you're on 100 or $200. And I think this is actually a problem. And I'll tell you why. You're going to learn to use these tools slower. So one of the things you'll notice as you get into this is that they're quite addictive and you can keep going in a session with Claude code for hours and hours on end or Codex for hours and hours **[38:19] Speaker A:** on end. But you prevent yourself from and in doing that, you also learn. But if you are hitting your limits consistently, you're just not going to learn as fast as you otherwise would. Now, if you've got $20 and that's it, I understand. Like you're going to get a lot of value just with the things I said in this talk with that $20. But you are going to learn slower. And so in like kind of what's turning into an industry like a race to figure out just like spend ridiculous amounts of tokens and uh know, it's a it's a risk that you fall behind peers who learn this faster cuz they just have bigger budgets, but then aren't able to figure out how to parallelize many aspects of work um better with AI. And so, I would think about how to reallocate towards a $200 a month budget. **[39:10] Speaker A:** So, one example of a grad student I talked to at Northwestern who's on the market this year got a great job. He got into this stuff early and he would just tutor on the side to have a bigger budget. And so, that's one way to do things uh if and then like if you have some research funds that you can push in one direction versus another, if your research budget is bigger than 200 a month, I would seriously recommend like increasing towards that amount. But I do understand the limitations of like you're in a developing country or you have a small budget, etc. Still, this is in my opinion a reality. And just one last point I wanted to make is that the job market for economists was terrible this year. So, this was in week 14 of each hiring season and the number of cumulative postings we saw, right? **[40:00] Speaker A:** So, it was down 31% versus the same point last year. And so, there's a lot of questions of why this happened. I think that a big reason about why this is happening is that the demand for economics undergrads will plummet. So, the types of jobs that an economics undergrad would do are these like kind of like white collar semi-technical jobs. And those are the kind that are being automated away with AI and to the extent that the demand for a PhD is tied to the demand for an economics undergrad, I don't think this bodes well for future job markets. And so I would heavily emphasize learning these tools simply like for your own interest, but like to make yourself someone who can, you know, weather the AI destruction that is to come. And so I I give these agentic coding trainings. If you found this useful, I do like group sessions and so there's a link here where I do like a more extensive, comprehensive session. **[40:58] Speaker A:** And also, if you just want to keep up with like what's going on in AI and economic space without having to spend all day on Twitter, you can let me do that for you. I keep pretty close to the frontier and so I've got a newsletter here where I sort of on a weekly basis approximately will send out a newsletter on just what's going on in AI and economics. So maybe of interest, but uh yeah, that's uh that's it. **[41:25] Speaker A:** Thanks so much, Aniket. That was amazing. Um we've had a load of questions through. I'm going to try and group them so we can get around to as many as possible. I guess a lot of them are talking about privacy and proprietary data. Um and asking about what precautions you would take when using these tools to work with kind of sensitive data. And um yeah, any any advice you have on kind of researchers who are using maybe individual-level data that that can't be shared or have signed signed various agreements to access the data. Uh how that factors into their decisions for using these AI agents. Yeah, um it's a good question. One is that like you could create a layer of separation between where the data is stored and executed and the writing of the code. And so, a type of thing you could do is to ask Cloud Code, for example, to create a profiling script. **[42:23] Speaker A:** And then, that helps it profile the data set. You just copy the results back without giving away anything that's at an individualized level or leaking PII. And then, either you could have this sort of back and forth manual relationship where you're never giving Cloud Code access to the code. It just creates a script and you execute it yourself. Or, there's two other strategies. One is that you can create a mock data set which has properties of the real data set. And then, apply whatever analysis you are trying to do to the mock data set. You can do certain kinds of anonymization, but this is going to depend on your privacy agreement. And then, use the real data, but like with the PII stripped. And then, the fourth option, um which I think is like it's difficult for an individual researcher to do, but it's perhaps feasible and reasonable for certain institutions. **[43:30] Speaker A:** Um but, is to get the infrastructure with which to have good local coding LLMs that you can use. So, these are large language models which um are completely local on your infrastructure. They're not going to Codex or Cloud Code's servers. And I guess, okay, a fifth option. So, that that that's expensive as a big upfront cost, but then that, you know, can be used indefinitely by your team. Uh the fifth option is that a lot of places with sensitive PII are able to work out agreements um, with places like, uh, Anthropic or OpenAI to have like a Claude or OpenAI instance that they're hitting for their Codex calls, but which is like private to their organization. And so, if you're able to work out one of these enterprise agreements, and it would fit under the auspices of your the types of data sharing agreements that you're seeing then that might be valuable. **[44:38] Speaker A:** That might be worth it for your team. But, that's like it's it's not a thing that an individual researcher would be able to do, but if you're at a place like Census or some government institution then that could be a route that works. **[44:52] Speaker A:** Um, another quick question on the kind of security side of things. Um, someone asked, um, if there's ways of ensuring that Claude Code or Codex does not install a package that may contain malware, for example. Uh, if you have any thoughts on that. Yeah, um so, one like one I I've I've never encountered that. Uh like you you might have heard of Open Claw, which is this kind of personal AI, which is a sort of wrapper around things like Claude Code and Codex, and then people were grabbing skills from a site called Claude Hub, and then some of those skills had malware. So, that that was like one attack vector. Um like like the if you're installing R or Python packages, then like the threat vector is going to be the safety of those reposit- safety of those package repositories. Um, another tool I use, which maybe I'll put on my site or I've talked about it on Twitter if you follow me there, is a tool called destructive command guard, which is a command line tool which just prevents **[46:11] Speaker A:** Codex or Claude code from doing certain kinds of destructive commands by default, like deleting a directory, just doesn't allow it. It can't even do it anymore. And so that's that's like handy. But then another like more advanced thing is that you can work with Claude code or Codex in a sandboxed environment. And so there's a computing technology called Docker, which allows you to create a sandboxed environment on your computer. And then if you work with Claude code or Codex in there, whatever {quote} damage it can do is limited to the sandbox. Um I think that Codex and Claude code are also making it easier to work in sandboxes. I don't right now. I just work on my computer. But a a final option, this is something I do do and I'm just going to start be doing more, is if you're really concerned about security, you can just work you can get a server, like an AWS server **[47:19] Speaker A:** or Digital Ocean or I use Hetzner. And you can do your work on this server instead. So like Claude code or Codex are just operating on a server that you SSH into. That's that's another option. **[47:36] Speaker A:** Um great. A couple other quick questions we've had. So someone asked if you if there are any best practices if you're working with coding agents in the context of kind of collaborating with co-authors who might also be using agents or might not be and you have these shared folders in Dropbox or or Google Drive. Yeah, let me um I'll uh uh I'll go to Twitter really fast cuz I I posted about this on Twitter just a little bit ago. So, it's right uh do do do **[48:15] Speaker A:** Ah, okay. Yeah, how should academics collaborate with colleagues who refuse to use Git and only want to use Dropbox? So, um I have a set of instructions here. Well, you got you can kind of read through it, but the long and short of it is that you basically will give Cloud Code if you if you're following my instructions, you should be using Git when you use Cloud Code or Codex, okay? And so, like that's the first thing is even though you're using Dropbox, you initialize it with a Git repo, but then also you just give it this set of instructions. Say, my colleague I'm collaborating with it refuses to use Git, which is a common problem, right? It's like a it's a matching problem. One wants to use Git, the other doesn't. We're working in a Dropbox folder. Can you update your agents.md/cloud.md so you're aware of that and so you can give me suggestions of changes you notice, which are likely colleague costs, so that we can handle committing for them. **[49:16] Speaker A:** And so, the problem which occurs when working with a colleague who has a Dropbox who doesn't want to use Git, let alone agenda coding tools, is that they're going to make all these changes that don't get tracked to the version control system. And so, it's like then annoying because it's incumbent on you to do it for them, but the way you can get around that is just by having the agenda coding tool be aware that this is a thing that could occur, hopefully identify them, and then do all the Git cleanup for you. So, that's that's my recommendation there. **[49:47] Speaker A:** Um a couple questions on kind of platforms and and types of types of agents. So, um someone asked what you think about the value of something like GitHub Copilot that gives you access to multiple coding agents at once. And someone else asked quickly um why you didn't mention Cursor. Is it basically just a a less preferable option for economists than than Claude Code? **[50:10] Speaker B:** So, uh GitHub Copilot is an atrocity. Like, it's it's absolutely horrible. Uh it just No, like, it gives you negative results. Uh I'm I'm not exaggerating, but it's just it's like known in the space as being an absolutely horrible coding agent. Uh having access to different models is not that much of an advantage because, like, Codex is the tool that makes the best use out of the Codex models. And Claude Code is the tool that makes the best use out of the Opus models. And so, like, if I want to use Opus, the only tool that I want to use Opus with is Claude Code and vice versa. Now, on Cursor, the reason why economists are drawn to Cursor is because they get um like an integrated development environment. So, I made a video here that nobody watched **[51:07] Speaker A:** on uh where is it? This one. Should economists use Claude Code, Codex, or Cursor? And so, I had a I had a different answer when I made this. I think I suggested Claude Code. And since since since then, my like default is Codex. But the reason I don't recommend Cursor is because Claude The main reason is that Claude Code and Codex are giving you a heavily subsidized rate in terms of the amount of token usage you get for the price you're paying. Whereas, Cursor doesn't get that subsidized rate because it's not owned by Anthropic or OpenAI. And so, the amount of token usage you get at $20 a a on Cursor is like nothing. And at even then at $60 a month, I think you'll get which is the second Cursor plan, you might still get more usage from Codex, and your results will be better. **[51:57] Speaker A:** The only demographic to whom I recommend um Cursor is uh front-end web designers. So, like on my team, I have a designer who's good at using Figma, a design tool, to make design changes to our application, and I want him to be more empowered to like collaborate with the rest of the team by making a change and going through all the way. And so, I'm having him learn Cursor, and I'm coaching him along in that, because Cursor has like a very nice sort of visual interface for web design specifically. For anything else, I don't recommend Cursor. **[52:36] Speaker A:** I'm just finding my notes. Right. Um sorry, I'm trying to keep up with got so many questions. Um So, I guess um a couple of quick ones for you. Someone asked how important, in your opinion, hardware is. Like, is there a recommended laptop you would go for that makes this possible, or if if that's not so much of a a worry right now? Uh it matters a little bit, because um I think like I'm not 100% sure about this, but each So, like let me pull up, okay? So, this is what the Codex desktop app looks like, and you see here, like I was working on this talk with Codex. And so, I mean, I'll just like skim through and show you kind of how I work on it. So, like Here, like I did a lot of research on Twitter to find like different tweets that would make sense. **[53:27] Speaker A:** And then I I said, you know, here, "Okay, cool. Now, for the next slide showing applications of cloud code, do we have any candidate tweets already in research?" And so, it gave me, you know, some candidate tweets, and then I chose the best ones, and it basically took screenshots, put them into the slides for me, and then put the link down below. So, I I never make these Beamer slides by like coding by hand anymore. What a nightmare. But, I I wanted to just bring them over here on the left-hand side. So, you can see I have like 15 different projects that I'm working on in some level of activity. And I think what's the case, I'm not 100% sure, but like each one of these threads that's active, like while it's active, it can take up to 2 GB of RAM. And so, if you're like kind of like working on one thing or just a few things at a time, I think it's fine. **[54:15] Speaker A:** But, my goal is to get to points, and this is more advanced stuff, where like the agents are autonomously choosing what next task to do, spinning up new threads like this, and then doing more work. You might get memory constrained once you start like parallelizing your work a bit more. In general, in like a modern computing environment, I would recommend having just 32 or 64 GB of RAM. Um things like GPUs and stuff like that only matter if you're doing like computationally intensive stuff that requires that. So, just have like a decent amount of RAM, is my recommendation. **[54:51] Speaker A:** Uh and we have a We We have some questions about where if you have any recommendations on where to like systematically find different skills that have been developed by others. Uh you gave some specific examples, but if there's a place where where you would go to get them. And then, another quick one is other than coding and research, are there any kind of other uses you'd recommend skills for for academics specifically? Yeah, sure. So, um I'll say two things about skills. One is a They're are They're They're very decentralized. There's not like a central location of skills. I don't know if there's super value in it. There's a fellow um I recommend. He's on the list of people I recommended following. Um Antonio Mele. He's at the London School for Economics. And he had a project. Where is it? This guy. So, open him up. **[55:47] Speaker A:** So, he had a project called let's see, like awesome econ stuff or something. **[56:01] Speaker A:** Yeah, this one. Awesome econ AI stuff. So, it looks like here he's like sort of centralizing things that are found elsewhere. Uh the one thing I'll say is uh don't get too reliant on just downloading the skill and using it because it's not that hard to make them and customize them to your own particular needs. And so, that's the way that I would recommend things, like like working with Beamer, right? Having your own particular styles could be quite useful. But if we go over to um Chris uh Chris Blattman's site, right? A lot of these skills are for project management. Like how do you keep track of all your meetings, right? So, I have a skill where in any meeting I have with a client or with my team, we have an AI note taker come in. Then I sync that transcript to a fixed location on my computer. **[56:57] Speaker A:** And anytime that I need to find something, I have another skill which can do research across all those meetings to see like when some particular thing was discussed. And so, that's something that could be useful to you all, but I would recommend skimming through like some of these two sources. And then I guess I can also like all my skills show up here in the uh desktop app. And so, um I worked on this project. I guess I should mention Econ Now. This is um my project. It's like a list of It's got a database of every job market candidate and their papers. And so, like, I had a skill here to be able to modify the information of one of these candidates, for example. So, that's like, uh, uh, that's that's something that you can use if you're a managing a website to be able to easily manage a website. **[57:51] Speaker A:** So, there there's there's tons of skills you can make, pretty much anything you could think of. So, I'd recommend just looking through some of those skill repositories I mentioned, and you should be able to find something good. **[58:06] Speaker A:** Oh, yeah. And I was just, uh, on the other anything else other than, um, kind of coding and research that you would recommend? Oh, sorry, did you answer that one? I'm trying Yeah, yeah, I did. I did. Okay, yeah, no worries. I'll I'll pose one other one quickly. Um, a lot of people, by the way, have been asking about the slides. I've made a copy of them available on, uh, Vox Dev at the link I've put in a couple times. So, if you just scroll down a bit, um, you'll be able to find them there. And then there's also a recording of this on on our YouTube, um, that will stay on our website and our YouTube, so everyone can stay up-to-date, um, there. Um, I think so, we've had a couple questions from students who get access to, um, can use Agent KI and Positron via GitHub Copilot Pro, um, cuz they have a free option for students. **[59:01] Speaker A:** Um, they were asking what Claude Code and Codex will add to that if, um, that they don't already have access to for research. Uh, if you're using GitHub Copilot, you're just using a much inferior tool. It's like, I guess, like, kind of Microsoft's general strategy is to like, I guess, give it to you for free and then hope that gets you into Microsoft products, but uh let me let me let me like just make my point on Copilot. So, one of the people I recommended is this fellow, Affine Theorem, um Kevin Bryan. And then uh let me just see what he has to say on Copilot. Uh **[59:54] Speaker A:** Yeah, oh he's uh yeah, he's **[59:57] Speaker B:** he's he's a similar Copilot hater as I am. It's just a really bad tool, and like I get being a student, and like not wanting to spend money if you don't need to. I would say that universities' AI spend is very inefficient. And like if you can't figure out how to get the uni to pay it for you, you should still get on the $20 a month Codex plan yourself, and just use it. Like I feel very strongly about this relative to working on the free one, because you have to think about your like future labor market outcomes in the sense in the broad sense, like how much ac- is econ academia going to be around and what form, with what demand. And then like in the private sector, you want to be skilled with the best tools there are, so like keep getting employed. **[1:00:57] Speaker B:** So, that's my advice. **[1:00:59] Speaker B:** First thing's first. Thanks so much, Aniket. We've we've come to the end of the hour, and there's many more questions than than we could get around to. Thanks so much, everyone, for joining. I mean, we had I think 750 people watching live between here and YouTube, so that's great. Clearly, there's a lot of interest. Um Yeah, I guess a simple takeaway from this is to subscribe to Aniket's YouTube channel. There's If you're watching on YouTube, there's a link in the in the description to to his channel. These slides are now available on VoxDev. I just put a link in the chat to a blog where I've been trying to compile everything I find that's useful for AI and and development economics. And yeah, thanks again Aniket for your for your presentation and sticking around to to answer so many questions. Yeah, absolutely. So, glad to be here. **[1:01:45] Speaker B:** Thanks. See you. Bye.