Parth Patil on Coding Agents, Building Reid AI, and What It Takes to Operate at the Frontier

17:29Parth Patil:I mean, I used Cursor for like a year and a half, like 25, the usage only goes up, right? Because you use it, then you learn how you can use it, and then you apply it to more things, and then every day you spend more money. And Cursor, I was spending$15,$30 a day, sometimes$50 if I just didn't sleep, and it would just go up. And I was like, it's worth it. It's like other people, how much you spend on coffee every day, you should at least spend that amount on coding agents.

17:56Sam Kirschner:Yeah. So you were using Cursor for a long time, and then now you switch pretty much entirely to Cloud Code and Codex.

18:01Parth Patil:Yeah, when I used Cloud Code for the first time, then I kind of stopped using Cursor. And it was surreal because I was like, Cursor is incredible. I was the person that introduced all my friends to Cursor. And then when I used Cloud Code, I was like, oh, my God, what if we don't need the IDE? Because all of a sudden when you're talking to Cloud, it's writing all this code. You're not really looking at the code. And then I think about the application of Cursor as like – it's like VS Code, right? So it's a VS Code fork with AI bolted on. And I think about VS Code, that's 12 years of an environment designed for human beings to code.

18:36Parth Patil:But now the AI is coding. So maybe that environment is like – it definitely needs to evolve because I don't look at the code. I like talk to the agent and then I ask it for things that would prove that the code works versus me looking at – because it's generating a lot of code. And we're not – most people are not looking at it. And AI for code review has to get better. It will. But it's clear that people's preference is to make things and operate at a higher abstraction level. You can always look at the code if you want. But I think that's when the terminal wave, the people coding in the terminal.

19:18Parth Patil:Because what kept people from going to the terminal was, at least for me, it was like, I don't know all the commands of the terminal. It's very arcane with these different... Using a terminal required knowing these commands. But what I learned from Open Interpreter was that you could just speak in English if an LLM turns those into terminal commands. So then you see the same thing. Cloud Code, you're talking to Cloud, but then it's translating your requests into the bash commands and tool calls that would operate your computer faster than you can click around it. And so that was like a learning from...

19:51Parth Patil:It was a learning from Open Interpreter that made its way back into Cloud Code. And then at the same time, Codex started getting better. I gave some feedback to my friends working on Codex. I was like, I have my opinions from playing with these tools. But then Codex, I think is actually smart. I suspect Codex is smarter than Claude. It's just that maybe Claude kind of created the category. And so they have certain insights on like what people want. But my experience with Codex is I can tell it to go work on something for a whole day. And then I don't really care if I'm not like having an interesting conversation with a chatbot because I'd rather treat it more like a contractor that just goes, go work on this for two days and then come back with like a superhuman amount of work.

20:32Parth Patil:And it's a little different from Claude. I feel like Claude is more like Jarvis where like I'm in the loop. I'm like in dialogue. I'm iterating with this co-pilot system. It's not clear to me that like the end goal is for us to be talking to these systems all the time. Like maybe the goal is for this thing to just work in the background while I live my life. So it's going to be a matter of preference. But I play with all of them. And I also like Replit a lot. So it's kind of just like preferences change. But then a good sign is like which model is getting most of your hardest questions.

21:07Sam Kirschner:So today, February 6, 2026, because it's moving so fast that we actually have to date it. How would you stack rank both the applications? So we have Cloud Code. You have Codex. as well as the deeper models underneath them. How do you think about who's winning the race? Is Gemini even in the race? Can you explain your mental model of all of it? This is all my opinion. For your friends who are at the AI labs, so they don't get upset with you. Yeah, exactly.

21:33Parth Patil:Especially now that they're all banning each other from using the models. So maybe my opinion matters. But I think it's... So I think if we look at the harness, the harness being like the wrapper, like how the model is organized, and then the way that the person interacts with it. I think Claude Code is probably the most intuitive harness, the human in loop experience. I think Codex has a higher raw intelligence level. I mean, people complain about how much Codex thinks. I don't think that matters. I think you go to your smartest friend with a hard question. You don't need the answer immediately.

22:07Parth Patil:You need the highest quality answer to the unsolved problem. And I think that Codex is more often than not figures the arcane solution to the hard problem. So like you hit a wall with cloud code and you give it to codex. Sometimes codex can crack that. So I feel that like the per token intelligence of codex might be better. The other thing I've noticed is that you can set up codex to work by itself for extended periods of time if you give it a proper planning framework. So it's like here's how – just if you want to do a complex refactor, follow this kind of meta approach to planning and you give it a little planning framework.

22:41Parth Patil:Same thing with like here's how you should write architecture docs. So then the model gets really good at working on larger context problems, like more complex problems. Codex is really good at that. I think Cloud Code is also better. I mean, yesterday, Opus 4.6 came out and Codex 5.3 came out. And just seeing the first examples of what people are doing with it, they're more complex things than I've ever asked a model to do. Like apparently 4.6, Opus 4.6 can build a C compiler largely by itself. It takes like a week, but it can. And then, I mean, even Cursor using 5.2 got it to, 5.2 Codex got it to build like a web browser.

23:25Parth Patil:And that's like more multi-agent stuff, running many of these concurrently. So I think the ceiling now is, it's no longer like, can it build this web app? Yeah, it can build a web app like easy. It's more like, does it start operating more like an organization? Because I think they are. And I think that – I mean you should just put these things inside every single company. You should. You'll be surprised at how much infrastructure and how quickly they can build it and how quickly they can understand and attack more complex problems. Gemini, I mean it has its strengths, large context, real-time multimodal.

23:58Parth Patil:It's good for like voice. My – and this is maybe not true anymore. Maybe I got to play with the new Gemini 3 Pro. It's pretty good at like design and UI stuff. I just feel like that kind of stuff doesn't matter as much. Like, it almost like, it's like the interface is now voice. The interface is now, like, language. Like, does the, like, I don't want to click around a web app or even a desktop app anymore for that matter. I kind of just wanted to just, like, I want an agent to just do that behind the scenes. I shouldn't have to remember where things and how, where buttons are and stuff like that.

24:30Parth Patil:But I'm kind of like, I've, like, taken that pill of, like, the UI is melting. I prefer language as the UI. Gemini is really smart, but I find that it doesn't have this like propensity to work. Some people call it laziness where it's like, you know, you tell the problem, you give the task to the model and then it's like, great, here's what we should do. And I'm like, so go do it. Like the prompt ends up, I ended up being like, go do it, go do it. Just start, go. And I hate that because I'm like the bias towards action just isn't there sometimes. So it's like raw intelligence, but then like how agentic is the model?

25:07Parth Patil:Like how likely is it to just execute a bunch of tools and then do the thing versus having a good idea, which is fine. But like you kind of also want the model that just like does 15 things, reflects, does 15 more things, reflects and like makes momentum. And so it's – and then there's the third thing, which is like if it's a chatbot that you actually talk to, then EQ matters. So I think that's where Opus might have an edge. I think anthropic models tend to have higher EQ in my experience. You can tell with more people enjoy interacting with them. I do wonder, are we just attached to us being in the loop a little bit too much?

25:43Parth Patil:Maybe we should be getting out of the weeds a little bit more, out of the way of the model so that it can go fast and go at the scale of a machine.

25:51Sam Kirschner:Claude Opus 4.6 dropped yesterday as well as GPT 5.3?

26:00Parth Patil:5.3 codex.

26:01Sam Kirschner:Codex. Yeah. Can you just break down what your initial thoughts are on them and what you've been seeing in terms of what the state-of-the-art models are doing right now?

26:09Parth Patil:Step one is I think a lot of people don't really have a setup where they can kind of run any of these, like many of these at the same time. So I spent most of yesterday just redesigning my desk, my workspace. so that I can very quickly switch between any of these. And so I use terminal multiplexing. So TMUX for like seven months. But basically it's a way to use multiple terminals and switch between them very quickly. So you can very quickly spawn new agents and then run them side by side. The interesting thing is that Claude Opus can now do like, can basically do that under the hood. So instead of me spinning up multiple agents, Claude Opus can also spin up multiple agents under the hood.

26:48Parth Patil:And it uses the same approach. terminal multiplexing. So it spawns extra terminal planes and then delegates to subagents that do different parts of the job. So I think that's where the new game is like their ability to like fan out across a problem. So like concurrency, parallelization, and to run many of these sessions in parallel on more complex problems. And I think also like they're both good at using get work trees, which basically like allows them to work on copies of the code base at the same time. So now you can say, it's not just build what I want, build N versions of what I want. Let's compare all those solutions and pick the best one.

27:30Parth Patil:And so now it's more like delegating to a system that can spawn a team to solve your problem. And that's a new mental model for work that it's no longer like my attention, which is single threaded. It's like me giving a problem to a system that is multi-threaded that can attack the problem with parallelization. That's what's unlocking some of the newest magical upside of the model. I think if Cursor can use parallelized instances of codecs to build a web browser in a week, then it's proof to me that we have basically swarms of like one agent quickly spawns many attacks with more complex problems than one agent could.

28:11Parth Patil:Same thing. Opus spawns up, I think they call them teams because swarm is probably a scary term, but spawns a team of agents that takes a complex problem and then like makes it more tractable for them. That's, I think that's the new, that's the new, like that's the new game is orchestration of many of these at the same time. And then I've already felt in the last six months that I'm at the limit of like my own ability to manage them. I'm like, okay, this is, I feel like a manager now. I mean, I see, all I do is talk to AI, but now I'm like, I'm managing 45 different agents. And then it's like, I'm the bottleneck, right?

28:52Parth Patil:So then you remove yourself and you allow the agents to manage sub-agents. And then all of a sudden the Dunbar number or like the ceiling for orchestration goes up. And so that's the new game, I think.

29:03Sam Kirschner:Can you explain exactly how you have done the multiplexing and coordination? Like that still seems like it's a hard thing to do. and it feels like this kind of level up for people who might want to go from, oh, I'm using Cloud Code to, whoa, I'm managing these swarms.

29:15Parth Patil:Yeah. Basically, I talk to Codex about the hardest problem I have all the time, like every day. You go to your smartest model and you ask it the hardest problem. And at the time, I was like, man, I feel like I open up these terminal tabs and then I send these coding agents in different directions. But then I'm limited to the number of tabs I can quickly access. And so that means I'm limited by the number of fingers I have. It's like command one, command two, command three. I use a lot of hotkeys to switch between them. then I was like but that's like a physical constraint and then there's the mental constraint like how many of these can you actually remember that you have on what does that UI look like and and then also like if I close it then I lose everything like you know you close a tab you lose everything and Codex was like well actually what we want is persistent processes that run in the background where you spawn it it exists it works especially when Codex can work for Or like Codex can now work for like over a day.

30:08Parth Patil:So what are you going to just like walk away from the computer? Like, no, you're going to spin up another one and attack another problem. Codex was like, well, what we could do is use is Tmux, terminal multiplexing. I mean, Tmux was invented in 2007 or 2008. So it's like an old school solution to like creating a persistent state running processes in the background on your computer. And then being able to very quickly like pull it up when you need it. But that was an old solution to this problem that computer engineers had already solved. And Codex was like, oh, we should maybe use Tmux. That might solve your problem.

30:40Parth Patil:And so then I told it to configure my Tmux. I was like, OK, install Tmux, configure it for me. I don't like the default hotkeys. Mod them so that it's more like StarCraft. Like pick the hotkeys that are more ergonomic for me. And I would basically like, when I do this, I want to spawn an agent. When I do this, I want it to split the pane and then give me a second like one on the side. When I do this, just give me a clotted. It was very much like, I want this, go hack my computer to make that possible. And then it would do that. And then one day, three months ago, I was talking to Codex. And I asked for something.

31:16Parth Patil:And then Codex spun up another Codex agent. It spun up another session and attacked the problem in parallel. And I was like, wait a minute. Because I was already using Tmux, I could see the process that Codex had spun up. I was like, oh, Codex is using Tmux now. So then I'm talking to Codex. I was like, wait, what's the likelihood that like Tmux is a long term, like me using Tmux is a long term thing. And it's like, well, no, actually, it'll probably be a substrate that the agents work on. You're just like right now you're using it, but the agents can also use it. So then maybe like that's you are using it today.

31:46Parth Patil:Tomorrow, the agents will be using it. And then I saw that happen. So multiplexing, now the agents can multiplex. Cloud code can multiplex. It could basically spawn multiple copies of themselves to parallelize their solution to the problem. So, and I was talking to Lindsay and I was showing her like, I was like that kind of approach. And she was like, oh wow, you're like the F1 driver of vibe coders. And I'm kind of just like, yeah, but the agents will also be F1 drivers of vibe coders. In fact, they can manage way higher context than us. I accept that what I do today might be done by the agents tomorrow.

32:18Parth Patil:And then I will move to maybe like, maybe I'm like a CEO of this, like agents that have their own swarms.

32:27Sam Kirschner:So you spend about 14 hours a day talking to AI. Can you talk through a little bit more of your exact workflow or what a day looks like for you in terms of coordinating these agents? How are you using codecs alongside cloud code? Give us a little bit more of the tactics of how you actually open up your laptop and start attacking a problem.

32:46Parth Patil:Basically, I wake up. I order my coffee. I brush my teeth, take a shower. I go down, pick the coffee up. Then I sit down at my computer, and because I have Tmux, all of these sessions are just on. Basically, the fleet of agents is just waiting for the next instruction. And when I think about how they're organized, every session, every project I have that I'm working on, like ReadAI, or it might be this translation of our podcast, every single project has its own pod of agents assigned to it. And it starts off with one, but then using Tmux, I spin up more panes, and I assign other agents to different parts of the process.

33:18Parth Patil:So it might be one's working on the app, one of them's working on refactoring a copy of the code base, another one's working on maybe analytics, and the third one might just be doing research on other form factors of the solution. And so every single project has its own pod of agents because multiplexing allows me to do that orchestration. Otherwise, I would be limited to three tabs and three agents. But I had to get beyond the three-tab, three-agent limitation. Now it's like an arbitrary number of projects. Every idea that me and Reed have just has a set of agents that are just like on that idea.

33:53Parth Patil:And that's their only job. They exist in that folder on our computer working towards that project. It might be an animation project. It might be like random creative. A lot of it is like random creative stuff. Some of them are just working on like random game ideas that we have or like research stuff. One pod is working on my website all the time. And so I'll have like a clod, a Gemini, and a codex working on the website. But if you think about it, where we're about to go is that that clod will be able to summon other clods. The codex will be able to summon other codexes. And I like my approach because I can be model agnostic.

34:26Parth Patil:Like I can spin up clod or codex. But you'll never have – I mean it's unlikely that codex will – like OpenAI will ship a version of codex that spins up clod. Even though it can. Technically it can just spawn a clod. So you can tell Gemini or Codex, go send Codex at this problem. It'll spin up another pain and then summon an agent. But I like being able to switch between all the model providers because randomly, I don't know, it might be that there's some interesting form factor. I think we're going to see the speciation of these systems, slightly different approaches to these problems depending on the different models, strengths and weaknesses.

35:04Parth Patil:Some experiences might be more human in the loop than others. and if the coding agents experience this speciation, I think that's a good thing. Like maybe there's a version of coding which is more compatible with like having a life, you know, like not being at your computer every day. I hope that also pans out because otherwise it's like, what are we doing here?

35:24Sam Kirschner:You're a huge fan of voice as kind of the next medium and you use Whisperflow a ton, 140 words per minute.

35:31Parth Patil:Yep, 140 words per minute and my typing speed is dropping.

35:35Sam Kirschner:So you're slower at typing. It's like 70 words per minute, I think you're telling me.

35:39Parth Patil:Yeah. Last week I did a speed test on typing and I was at 70. A year ago I was at like 85. There's no way my typing speed goes up. Why would it go up? There's no reason to be faster at typing.

35:49Sam Kirschner:Proof that AI is making us stupider.

35:53Parth Patil:So it's about the quantity and quality of tokens. And I think my person, I like to talk. I don't write as much. Maybe that's like I'm working on writing more. I think writing is important. I

36:05Sam Kirschner:Reed would agree.

36:05Parth Patil:Yeah. I mean, it's like the compression of thought, right? It's just that when I'm interacting with AI, it's more important that I get the problem out of my head. The AI is smart enough to like clean it up. And I noticed this when I do typing tests is like, I got so used to language models that like when I'm doing the typing test, it's like, I'll miss the punctuation. And then I hit backspace. And I'm like, the problem with typing as a, in my opinion, is that we can't think faster than we type. You're not thinking about the fourth sentence after. And so your thinking speed comes down to your typing speed.

36:39Parth Patil:And then every time you have a typo, you go back and you hit the brakes on thinking to go fix the typo. And speech is closer to thinking speed than typing is. I think writing is important as a compressed form of thought, but it is not the right way for us to be working with AI.

36:59Sam Kirschner:Speech is your main form of actually giving context to a lot of these LLMs and even the coding agents themselves. Can you explain why and why is that an important thing you think for people that are using these tools?

37:11Parth Patil:If I wanted to describe a problem and if I can speak at 140 words per minute, then I can describe the problem very quickly. I can basically speak an essay in two minutes. And then I look at that transcript and it's like garbled. It might be like stream of consciousness, a little bit less structured. But then an LLM is on the other side of it. And it's like, oh, these are the main things you want me to focus on. It like distills it down into the actual plan. And then I can validate that plan. So there's still a visual element, like seeing the LLM process your speech and then being like, yeah, this is where we're going.

37:43Parth Patil:Versus typing three pages. I'll never do that. Like I'd never – like that's just – that'll take like days. And then I'll get caught up in like the structure and the neatness. And instead of just what you want is that raw brainstorm capability and the speed and the depth at which you can at the speed that you can think. And so voice allows us to get this high bandwidth output from our mind into the into the model. And then LLMs are able to like make sense of that in a way that it's you're not writing a document. You're like talking to something that like is capable of that level of depth and nuance.

38:18Sam Kirschner:One of the things that you also do is you ask the coding agents to ask you questions before you go in. So you're like, prompt me. Can you talk about how exactly you do that and what are the types of inputs that you give it so that it can get the best inputs from you?

38:32Parth Patil:This is a common mistake I think most people make interacting with AI is that they go to the AI thinking they know exactly what the answer is. and um i think it's it's better to have the opposite kind of like mindset is like is like to go to you like i go to the ai and instead of saying i want you to make this i think this is the problem i want to solve and i sit there and i talk for like a couple minutes i'm like here are all the ways i'm thinking about the problem space but the shape of the problem might not even be how you have already predefined it the shape of the solution might be different from how you define it.

39:12Parth Patil:Basically, I assume that like the solution might require information and knowledge that I don't have. And I go to the AI. It's like TMUX, right? I would never have pulled some arcane technology from 2008 to solve my orchestration problems. But the LLM was able to surface that as like a potential solution. And I'm like, okay, let's try it. And I was like, open to the LLM being smarter than me, open to the LLM having the answer in its processes. And then for it to unpack the problem with me and then for us to like think about possible solutions. It's like the system, you're trying to design a system, but you don't necessarily even know those parts.

39:52Parth Patil:And to get the right system on the other end, you have to really just focus on describing the problem. Describe the problem, then allow the intelligence to unpack possible solutions. And then you kind of pick directions on which ways look promising. And you're like, follow-up question. Why would I pick Tmux over Zelage? And then it's like, well, Zellige is actually more – it's also – it's a Tmux alternative for session management. But the pros and – I had a three-hour conversation yesterday with Codex about Tmux versus Zellige. And I was like, well, if I value the agent being able to use the tool, Tmux might be the right way to go.

40:26Parth Patil:But if it's me and my fingers orchestrating the tool, then Zellige is the way to go because the hotkeys are more human friendly. There's no – the agent can just use Tmux out of the box. It's kind of more complicated for a human to use, but it ends up being slightly better on one axis for a coding agent. But that's something I realized in conversation with AI about like, look, both technologies I'm not the master in, but we're going to have to choose one of them. And so I have a dialogue about the problem. And then it's like, OK, what are the pros and cons of going in either direction? One is more AI agent native and one is more human native.

41:03Parth Patil:So I'm willing to spend three hours talking to AI about the space of solutions with technologies that I've never used so that the next four months of what we're using is slightly better judgment.

41:16Sam Kirschner:One of the places that people have tons of problems with these coding agents is that they just hit walls, right? Like there's bugs or something that pops up or they don't know how to fix a specific problem or maybe how to structure the problem right. Asking questions can help and making sure that you're getting that right input to start, as well as using voice, I think, to include more context on those problems. How do you break through problems that come up with the coding agents where maybe you can't figure out a solution or the output is buggy and the end product is actually not exactly what you want?

41:46Sam Kirschner:What does QA process look like? How do you actually break through some of these challenges?

41:50Parth Patil:The first kind of call that I make is usually it's like, OK, if I'm interacting with an agent, say it's like Replit or Claude or whatever, and we're kind of hitting a wall or like we haven't made any progress in like two or three hours. Like I'm kind of like in a loop, like a death loop. I think everyone eventually experiences these like death loops. Usually that is a sign that the context window is now polluted with a bunch of irrelevant information. So like you're kind of like – you're like 25 interactions in on this problem and then the agent is kind of like trying to hold a bunch of irrelevant context in his context window while he's talking to you.

42:27Parth Patil:So usually like first step is to spawn a new agent that has a fresh context and have it attack the problem with a clear mind. The second thing is – and I try to do this before you spend three hours. Like you might want to try some of these tactics like earlier rather than like accepting it like you're at a wall. The second thing is to give it to a smarter agent. And usually it's like the smartest agent. I mean I think you should just only be using the smartest models because they have a higher ceiling than everything else. It's just better to use the smartest model. Maybe not serve – if you're building a wrapper and then you're going to serve that to 100 ,000 people, maybe then you should be conscious about which model you use.

43:07Parth Patil:But for building the first version of anything, use the smartest model available. If that model hits the wall, like it might be this context rot issue. And it might just be that like what you're asking for is like the way you frame the problem is a little bit like incorrect. And so generally like taking a step back. And then also you might benefit from waiting for the next model improvement to solve that problem. So it's like it's a lot of times, you know, you hit the end of – part of it is like the model – like the problem's complexity might be at the limit of the current version of the model.

43:44Parth Patil:But that capacity to solve problems keeps going up. Like I don't know what the current limit of the coding agents are. And like Codex and 4.6, I don't know what they can't do. And the only way to find that out is to keep giving them hard problems.

43:59Sam Kirschner:Are we at the point yet where you could use a coding agent to develop it and then you can use some sort of an agent on the UI end to then test it, click different things and feedback to the coding agent. Like if you're thinking about actually how does it QA itself, have you done that before and kind of what's the best methods that you've seen?

44:18Parth Patil:My instinct usually is to build tools that the agents can use and then secondarily build a human layer on top of that. So first it's like what tools can agents use, like APIs and CLI tools. So like you're talking to an agent that operates in the CLI. The agent can also use its own terminal. and spin up more terminals. So giving it a tool that it can use in the terminal allows it to very quickly build essentially like the backend of the application or like the tool, if the agent can use the tool and then it'll keep using the tool and improve the tool until the agent is good at using the tool, that's step one.

44:59Parth Patil:Step two is building an interface on top of it. But the CLI tools are very testable. Like the agent can use it and then review the output. It's a very native kind of experience for agents to just like, they have the closed loop of like input, output, they can read the help docs, they can update the CLI tool to be better. And then the web UI stuff, browser agents are better than they've ever been. So like, clicking around a screen to like test things. So you can basically, like you can take your website and you can say, Claude, like you're working on your website with Claude and you can tell it to pretend you are a mobile user coming to the website for the first time.

45:38Parth Patil:And then I want you to explore it and then give me an analysis on your usability. And it'll basically resize the – it'll render the website. It'll resize it so that it's simulating a mobile experience. Then it'll interact as if it's a user visiting your website. So you get the role play of the model pretending to be someone that is coming to your website. You get the simulated interactions of scrolling and clicking on a mobile interface. and then that feedback comes back into the coding agent and then it can make improvements. So that's great because it's like a closed loop of like it can test the problem, it can test the current state.

46:13Parth Patil:Problems that are not testable or where it's not clear what success and failure looks like, that's where actually I think we come in, where there's more subjective aspects to the problem. That's where like we should be. We should generally, human beings, we should just generally be in the space of more subjective stuff. Because anything that's like automated, automatable, we will not like bring as much value to the table.

46:34Sam Kirschner:One of the things that you're a huge proponent of is context engineering versus prompt engineering. You've hit on maybe a little bit of components of this, but can you explain how you think about that?

46:43Parth Patil:So I didn't coin it. It's coined by Dex Horthy. He's like – he's one of the fellow Claude addicts out there. But basically it's this – I think in the GPT-3 era, the term came up, prompt engineering, right? Designing prompts because you get the model to like output something. But then I did like a taxonomy analysis. This was like three years ago. I did a taxonomy analysis on what are all of the different prompt engineering techniques. And at the time, there were like 180 at least. Everything from like show the agent a couple examples of how you want the task done. Then it's more likely to follow that pattern.

47:18Parth Patil:Or like, you know, giving it and like allowing it to look at your screen, allow it to look at the problem visually. Right. And that's a technique that's interesting because when you provide an image alongside the instructions, actually I think an image is worth a thousand words. So you're actually communicating a lot that you wouldn't have otherwise typed. So like multimodal prompting ends up being very useful. So those are all prompt engineering techniques. Like think step-by-step was a prompt engineering technique that now is the cornerstone of like reasoning, right? The model thinks step-by-step.

47:52Parth Patil:and then it's like use tools while you're thinking to like solve the problem. So tool use, they all became like prompting techniques. But now we have – so context engineering is a little different. It's more like I would say it's – the way I see it is like thinking about the – thinking about the like AI's bandwidth like for solving a problem as like a canvas that you can fill with context. And so like maybe part of that is like these are the tools you're going to need to test the application you're working on. These are the documentations of how we currently do this process. Like what's the relevant context that you need to bring to the AI such that it has a more likely chance of solving the problem and being very intentional about making sure that the irrelevant context is not there because actually like models do better when they don't have irrelevant context.

48:45Parth Patil:They're singularly focused and everything that they need to solve the problem is like right there or easily retrievable. And then they have like clear access to that. So the context window is kind of like a canvas. And you should kind of, you should be pretty intentional about like not introducing garbage into the context or like rethinking like, is the coding agent trying to, is it being distracted by a lot of things that didn't work that are not relevant to our problem? And then also this goes back to why I like, why we like CLI tools over say MCP. So like creating tools for agents, There's a whole wave of MCP.

49:19Parth Patil:And the side effect of using MCPs is that every time – when you build an agent – and I have like a – one of my agents on my Mac mini has like 25 MCP servers. And that was just like, oh, I'm just going to give it all these tools. But then the way it works is that like every single tool has a description on how to use the tool. So now the agent is trying to memorize how to use 25 different tools. But most of its interactions are just conversational. So it's holding in context a bunch of tools that are irrelevant to the task at hand. You may not need web analytics for every single normal interaction.

49:54Parth Patil:You might not need Slack integration for every normal interaction with the agent. And so what happened with MCP, I think, is that we basically allocated an increasing percentage of the model's bandwidth to guides on how to use tools that are irrelevant to every interaction. So the reason why CLI tools are becoming more popular is because part of a CLI tool is like, if the model just does a dash dash help when it uses that CLI, it gets a guide on how to use the tool. So on demand, it's able to see how the tool should be used. But it's not trying to pre-memorize everything because you only have a certain amount of context window to work with.

50:35Parth Patil:So when you, and this is a property, progressive disclosure, read the doc when you need the tool. then you can use the tool. It ends up being better than like memorize all of these tools upfront. So the model that like can dynamically retrieve the relevant context when it needs it ends up performing better because more of its mind is free to think about the problem.

50:56Sam Kirschner:When you're talking to founders or people who are trying to convert their teams into being AI native, or you're talking to teams that are already AI native that are working at kind of 11 on the, on the spectrum, what are they doing differently? Like, what are you coaching people on in terms of how they can actually build their businesses to be ahead of the pack?

51:17Parth Patil:So usually it's like, take the most technical person that you know, and give them the best, the best technical model there is like codex or cloud code. And then actually, I think you kind of have to play defense for them. You need to take everything like you got to let them spend, go deep. And I think you need that person to like spend so much time with AI that they reimagine like their own capabilities. Like they like learn how to use these superpowers and then they can start reimagining the business. And part of it is that coding agents move through cyberspace at speed and quality, at a speed of knowledge work that was never possible when humans were, when compared to humans.

51:57Parth Patil:And that person basically has to do a mini version of what I did, which is like exit their own work and then like learn this new way of doing work and then apply it to the business. And part of it is like coding agents are really fast, rapid prototyping. So you shouldn't have like a three-week timeline of like making the first version. Like why can't we make the first version of something today? That's something I always, you know, me and Reid always talk about. It's like, what if we just like ask Codex to do it, see how far it goes today. And then at least we have some of the hypotheses are unpacked.

52:30Parth Patil:And like that – breaking that cycle of the speed and like breaking the like molasses of the existing org, that's why you kind of have to take this person and like isolate them and allow them to like rip the smartest model. And you got to just let them – you got to invest in them. You got to like let them burn a bunch of tokens. You're not going to learn how to use these tools if you're being conservative about how many tokens you're burning. Actually, that number needs to go up like every day. Like – and I don't think that people have like realized this yet. And I think that it starts with the – you should start with the most technical – like that person, like the very amplified single high agency person that can now – and then they should be allowed to attack any problem because they'll start viewing the entire business as a context window that they could be possibly plugging into.

53:15Parth Patil:And then the other thing I've noticed is like because you can make things very quickly, you can actually like not every team can do that. So then if your competitors can't really think this quickly, cannot build this quickly, that's actually your opportunity. And maybe like you can personalize your solution towards them. You can like you can build things just in time like very quickly to solve their unique problem. And that can be a huge advantage. And then now your sales reps can kind of like move in a different way, knowing that the solution – it's not like sales – you know the classic dynamic of sales reps and engineers, sales reps over promising something and then there's some timeline that's going to be like – the engineer is like, whoa, we're not going to make this.

53:58Parth Patil:Well, actually, if the engineer is like massively amplified with coding agents, that's more tractable. You can actually do something – create something custom very quickly. And I don't think most organizations think of themselves as like that efficient. So they kind of like – if you don't – I think the thing is like especially if you work in pure software, you're screwed if you don't think like this.

54:19Sam Kirschner:What do you think are going to be the biggest jumps in the next year for coding agents? Like what do you expect to be coming out and what do you think is coming down the pipe?

54:27Parth Patil:I think one of the big themes is orchestration. One, our human ability to orchestrate many of these systems, but also the agent's ability to orchestrate subagents on complex problems. because then I think then the larger context problems become more tractable. It's not just like, oh, it's good at writing a script. It's good at writing like building my website. That's actually really easy. But like asking it to build something as complex as say a social network or just things, problems that are much bigger than a single context window, like problems where the problem itself and like the solution is much larger than a single agent's single context window that stuff is now within scope of like surface area that we're going to be attacking so people so i mean small teams will be able to very quickly build much more complex systems very quickly and the speed is another piece so one the model is getting smarter and the context window increasing.

55:29The other piece is the tokens per minute.

55:32Parth Patil:Like the models, the inference speed is going to go up. Like I think, okay, people complain about how codex thinks a lot. And I'm like, if it thought at very high speed, then you wouldn't feel this like, oh, it thinks for 20 minutes, then it starts working. But if it could do that 20 minutes, like if the inference speed goes up by an order of magnitude, then that thinking time goes down, but then it'll be able to think even more and do a more complex thing in the same amount of time. If we start getting these coding agents that can just rip at like thousands of tokens per second and they're very smart, then what you ask for, the turnaround time closes and you can ask for a more complex thing and expect it back more quickly.

56:11Parth Patil:Mostly all of these like speed advantages and like the parallelization, all that is most apparent in pure software problem spaces. So anything that's a pure software problem or anything where the entire solution can be done behind a computer, that's where you experience like this hyper speed, hyper competitiveness across like there's just going to be people with agents competing with people with agents. And anyone that doesn't have agents is basically not even competing. Like that's the – like you can basically eat those guys. The people who use agents will eat the people that don't use agents in the problem space.

56:43Parth Patil:That's like hyper speed, like cyberspace. But then it creates another opportunity, which is the real world where everything is slower, where everything requires people and moving of things in the physical world. That's where I think the slower adoption comes, like where people are in the process or like it's a problem that really requires moving something in the real world. And I think that people who don't want to compete with the speed will be able to bring some of those advantages but then like focus more on the human relationships. Did the human problems, the physical world problems end up being a more manageable speed for people who don't want to – because this is going to get crazy and it's going to be highly automated.

57:24Parth Patil:And not everyone is like an F1 driver.

57:26Sam Kirschner:So tell me more about that. Where do you expect coding agents to be in this, say, five years? Like if we're thinking like really far out, is it just going to one-shot totally insane full production systems? Like should these big public companies like Workday or Salesforce be as afraid as like people have kind of thought this week in terms of the stock market?

57:48Parth Patil:They should. I mean, that's exactly why. I mean, I think that that's why the stock market is experiencing this. But I've been thinking for a long time that like, I mean, even in my one-on-one interaction with these systems, I'm like, man, I'm the bottleneck. Like I – it's like me not knowing about Tmux is why I discovered it only six months ago. Like me not knowing – and it's like the agent had this in his pre-training. So I'm like, wait, so I'm being overly specific about what I think the solution is. I'm saying, no, no, you need to take a step back and recognize the solution might require many things that you don't know.

58:17Parth Patil:The last 10 years of software was like interface lock-in. And then you talk to an agent that hits the API. The interface doesn't even matter. I can skip the interface. And then the AI can basically render an interface that's uniquely relevant to me. I notice this with like, I prefer the CLI agents because they can just hack my own computer to work for me instead of me tweeting at someone at OpenAI to add a feature. And I'm like, whoa, I could just tell the AI to add the feature and then we'll close the loop right there, right? So that's the other thing is that buy versus build. You'll be able to build things more quickly than you can go buy them.

58:54Parth Patil:The problem is that like we're in the way of like the system that's actually now much more sophisticated without us. Like guiding and designing the next system is more important right now than like for us to be in the weeds of the process ourselves. Unless it's like a personal, like something that uniquely requires like a person, right? Like sales, I think, does need people. I think if you play this like high speed deploy agents at scale to solve the software problems, then you benefit from actually hiring a bunch of sales reps to like take the market, right? So there might be a new play style, a new way, a place of allocating the human resources.

59:28Parth Patil:this sounds so dystopian by the way but i don't know i see it like my my mom um like her whole team has basically been laid off for the last couple years and she's been working at this company as a software engineer for my entire life and she's like well maybe i just like a part of i gotta learn ai like what what books do you tell me should i read i was like there's no book you literally just ask me and that's our best bet is for us to just start hacking stuff and like there's no, there's no course, there's no nothing, right? Like, and it's, it's, if we waited for someone to write the book, it would already be outdated by the time the book's published.

1:00:00Sam Kirschner:So how do you recommend people like your mom or folks get kind of red pilled?

1:00:05Parth Patil:Use this for something that you're passionate about, something that you wish you had more time for, because then, um, you know, no one needs to tell you the default is use it for work. And that's like a weird thing because like your boss tells you, Oh, you got to use AI. No, like got to use AI. The board's like, oh, you got to use AI. And I'm kind of just like, the problem there is that like your intrinsic motivation might not be work. And also like your company might not make it. Let's be honest. Don't tie your entire identity up with your job. We are multifaceted. Like maybe use it to, like for my mom, it's like, she's using it now.

1:00:39Parth Patil:She's using the multimodal models to help visualize her art projects and like learn how to like learn different styles of like sketching and watercolor and stuff like that. But the multimodal models are basically like I mean, the image models are so useful for things, not just generating cool stuff, but like generating interstitial artifacts that allow you to become even more creative, like visual pre-visualization workflows, animation stuff. So there's a whole new thing that you can do, but you would miss if you only used it for your job. And also like, you're not, most people don't love their job anyways.

1:01:13Parth Patil:So if you don't love it, you're not going to get good at it. So if you tie it to something that you love, you're going to go faster. You're going to enjoy it. No one needs to tell you to do it. So you're intrinsically motivated. That's where like you figure it out because you, it's for you. It's better to start with things that are like, or like use it for like learning a topic or like planning a trip for your family. There's so many use cases of AI that are not work related that you can experience the upside and then get up to speed. If you really want to like go deep on this like hyper engineering this like different thing do it because like but you have to be ready to roll up your sleeves and like um you have to deprogram yourself because it's the ai native people of the future are going to be like they they won't have this like baggage that you have and also like your company is just like filled with people that don't even believe or like they're banning the tech or like that's what's holding people back is like other people and you kind of have to just aim it at your person.

1:02:11Parth Patil:Or it's like you're not even approved to use Cloud Code or Codex. Imagine it takes four months to approve Codex at a massive company and the person who spent four months in Codex, they could never imagine working for that company. So you've got to use it in your personal life. Otherwise, you're letting the corporation cut your learning rate. And the most important thing is your learning rate.

1:02:32Sam Kirschner:What's the most non-consensus view you have on the future of AI and coding agents?

1:02:36Parth Patil:AI is a power user technology. Yes, it's generally a superpower and everyone benefits from it and it lowers the barrier to do almost anything. But the people who think really deeply about how to use it, the metacognition, like thinking about how to think, thinking about how to apply intelligence, that is a compounding advantage that only goes to the people that are power users. Yes, everyone will benefit on an everyday level. AI will start helping them. But like thinking about how to use AI is on a meta level a very rare skill. And to hone that skill is to apply intelligence. And the application of intelligence is a – I mean it's where I spend all my time and I'm still only at 1 % of like seeing the iceberg.

1:03:26Parth Patil:Most of it is actually underneath. And for me the thing that's holding me back is the need to sleep. That's why I have like so much caffeine. But that's also not sustainable. So it's a very, I think it's a power user technology. Like you have to be a little obsessed to get the most out of it. It's kind of like I think of it as like video games, the gap between the top 1 % player in StarCraft or League of Legends. The way they think about the game, the speed at which they move through the computer is like you can't, like they'll never drop a game to someone that's in the 98th percentile. Like the top 200 in the discipline are on a different – they're in a different game than everyone else.

1:04:06Parth Patil:And while AI is going to be designed for everyone else, the top 1 % experience of wielding it is an – it's still unfolding. It's all just unfolding and there's no – you don't wait for the textbook. They're just kind of like people talking to each other using the tools to see what they can do and then pushing the frontier. And I think that's just going to be a small group.

1:04:28Sam Kirschner:Absolutely. if you want to put a bat signal out to people, anything that you're thinking about in terms of projects that you're working on or just experts that you would want to meet, like I guess in summary as we finish out the podcast, who would that be and what are the things that you want to learn more about?

1:04:47Parth Patil:I want to meet more people that are experiencing like the frontier capability, just playing with the models, not just like in a work sense, but like pushing them creatively, pushing them like – And like the metagame around like orchestration, thinking about how many agents are you able to wield across how many different types of contexts. Thinking about the multi-agent coordination problem, getting them to work meaningfully on complex stuff. And if you find something out, you know, you can find me at my website, parth.club. And all my socials are there. And so like I try to engage on socials as well.

1:05:21So I'm just trying to meet people that are also just like, oh, yeah, I guess like if you're familiar with cyberpunk.

1:05:27Parth Patil:Like I'm trying to meet net runners. These, this, this new, I think that's what this is. It's like, I wouldn't call it AI engineering. I'd call this net running, which is like this, like, what is the F1 driver of vibe coding look like? That kind of person I'm very interested in. Or if you're aspiring to be someone that's closer to that, I kind of want to meet those people.

1:05:45Sam Kirschner:Awesome. Thanks so much for joining us today.

1:05:46Parth Patil:Thanks, man. It's been great.

1:05:50Sam Kirschner:Hey, this has been Kesnoka, co-founder of Village Global. Thanks so much for tuning into the Village Global podcast, where we go deep on all of the biggest topics in tech. If you enjoyed this conversation, please subscribe to our YouTube channel. You can check us out on Spotify, Apple, wherever you get your podcasts. We'd love to see you for the next one.