AI Agents (i.e., Claude Code or Codex CLI) are the most important AI development in the last one year. And in the last couple of months, they’ve become incredibly powerful. In this article, I’m hoping to give you a visceral feel for the amazing things they’re capable of.
The Situation Until Last Month
In March, I said Start Using Claude Code or Codex CLI. “The learning curve and the cost is well worth the trouble”, I said, and “using a desktop AI agent like Claude Code or Codex CLI will change your life.”
Claude Code is a desktop AI agent from Anthropic. The equivalent offering from OpenAI/ChatGPT is called Codex CLI. The basic idea is that it is a downloadable program that you install on your laptop, and then you get to run Claude or ChatGPT on your data, on your emails and spreadsheets and documents and powerpoints, and you can leave it running as it does analysis, creates and runs and tests programs for you, and downloads and installs software for you, and generally becomes an assistant for you to do all kinds of things that weren’t possible for you before.
I gave a bunch of examples of how I and other people were using agents.
In July, I said:
Among those learning AI, there are 3 kinds of people:
No agent: Those not using agents (like Claude Code or Codex CLI) at all for their regular work. (They might still be using the web based interface for some research, or writing reports or emails)
Single agent: Those using a single agent, synchronously, throughout their work day
Multiple agents: Those using 2-4 agents without watching
And I exhorted you to be at least at level #2 if not level #3.
Examples of How I’m Using Agents Today
A couple of weeks back, I decided to become more ambitious with my own uses of agents, and here is just a few selected examples of what I’ve done since then.
Online Learning Platform
I’m involved in a project that will run a 5 month online course on AI awareness and education for undergraduate students. We expect it to have 5000+ students. The overall system needs to be very scalable, capable of handling large bursts of traffic when a new lesson is released to the students, but also able to shut down unnecessary servers and services at other times, to cut costs. It needs to be highly available—meaning that the system should continue running no matter what happens: it should detect any crash or error in any component and should automatically repair itself by deploy alternatives and backups. A system like this involves a fairly complex architecture and configuration in the cloud, including:
A pool of servers which are started up on demand and shut down when the demand goes awa
A load balancer which automatically ensures that none of the servers gets overloaded
A sophisticated monitoring system which looks for any errors and remedies them
A different sophisticated system which monitors the platform for any potential cyber security attacks or other suspicious behaviour
A third system to monitor the costs and send alerts if anything is going over budget
An email system that has to jump through various hoops to ensure that the system emails don’t get marked as junk mail
And many other things (Codex CLI made a list of 19 items which I’m not reproducing here)
A friend of mine who runs a software services company tells me that this would have taken a team of engineers a couple of months to build.
Claude Code and Codex CLI built this for me in 2 days. The video above tries to give you an idea of this project.
And, by now, you should have guessed that the entire video was also created by the same agents.
A board game
My family and I are big fans of board games. Our games tend to be quite complex—and just explaining the rules of a game can take over half an hour. Those of you who’ve played Settlers of Catan would know the kinds of board games I’m referring to (Catan is one of the simpler ones).
India Rails is one of our games. And now that our kids are living in different cities, I thought it would be nice to have an online version of India Rails. Except, India Rails doesn’t have an online version. So, in March, I decided to check if the agents could create one for me.
That project was an utter failure. The agents were terrible at understanding the two-dimensional nature of the board in the game.
In July, when Claude Opus 5 came out, I decided to give it another shot, and overnight, Claude Code with Opus 5 recreated the entire game for me, based on photographs of the board I gave it (and rules that it downloaded from the website).
Check out the video above (once again generated by the agents)
Editing the AI Vey Podcast
Last month, Ashish and I also launched AI Vey, a podcast about AI in India. What I didn’t realize until last week, is that Ashish, who is an economist and does not have a software background, is in some ways an even more accomplished user of Claude Code than I am. He showed me how the entire workflow for editing an episode of the podcast can be done by agents, and now I am also doing the same.
Specifically, these are the steps:
Log into riverside.com, he online platform where we record the episodes, and find the episode that needs to be edited
Look through the transcript and decide which parts of the recording need to be deleted. This includes the general chit-chat at the beginning before the introductions, interruptions during the recording which need to be removed, long pauses and silences that can be removed, and other cuts.
Adjusting the audio levels of the 3 different speakers so that they all are equally loud
Adding the standard intro at the beginning of the podcast, and the standard outro at the end
Automatically generating the shownotes, including a high-level summary of the episode, links to all the references mentioned in the episode, and more
Suggesting what short clips can be extracted from the episode to post on social media. This includes creating clips that involve picking up related segments from different parts of the episode and stitching them into one clip. And it also involves adding subtitles to the clips.
Of course, it invites us (Ashish or me, whoever happens to be working on editing that particular episode) from time to time to doublecheck things and make sure that that everything is in order. Which we do.
Check out the latest episode, There Is No Such Thing as Sovereign AI — Pranay Kotasthane, which was fully generated this way.
TMKK
So what should you do about this. I don’t mean to imply that all of you should now deploy a scalable cloud platform, or develop an online game, or start editing your own podcast.
The intent of this article is to give you a feel for how sophisticated the agents have become and the variety of things they can do. At this point, if you are unable to use agents to automate large chunks of the job you used to do last year, you need to introspect.
And if the reason for not doing this is that you don’t feel justified in spending ₹2000 per month on ChatGPT or Claude, then you need to do a better job of calculating investment and returns on investment. (How much did you spend on your college education?)
And if you’re a decision maker in a company, and you’re not pushing for this in your company, well… you and your company need to start worrying about competition from companies who are doing this.