Pi is the Claude Code Killer Nobody Saw Coming...

[00:00:00]A lot of people have tunnel vision when it comes to AI models and treating them as if they are the entire product, but it's missing a big part of the picture.

[00:00:07]If you take the exact same model, the result can change dramatically depending on what coding agent you're using. So, it's not just the model, it's the coding harness wrapped around it. And there are two big competing philosophies for how these should be built. First off, we have coding agents like Claude Code and Codex, and these can be referred to as having batteries included. They have huge amounts of features out of the box, a big bundle of tools, very long system prompts, and their own opinionated ways of doing almost everything. And that's why they're so easy to pick up and start using. But that convenience comes with real tradeoffs. They're designed for a wide audience, and features are getting more and more bloated over time. They're also a black box, so we don't really understand what's happening under the hood, and generally only really work best with expensive frontier models. On the other hand, there's a huge counter-movement pushing back against all that. The coding agent Py has been gaining a huge amount of traction, and it's deliberately stripped down to almost nothing by design. We'll be digging into Py a bit more in a few minutes. But first, let's have a look at how combinations of coding agents and models can really impact things. Here we have the Terminal Bench 2.1 leaderboard, and we see how different models perform with different coding agents wrapped around them. And you see some pretty significant differences. For Anthropic and OpenAI, their own harnesses win here. But you see the very minimal Terminus 2 coding harness, which is a harness just really used for benchmarking, performing nearly as well in some cases. And in fact, with Gemini 3 Pro here, it even outperforms the Gemini CLI itself. So, what even is a coding harness? At a very high level, think of the model as an engine, and the harness is everything else in the car.

[00:01:44]The engine gives you the raw power, but whether the power actually gets you anywhere depends on everything else in the car, the steering, the gearbox, the brakes, and all the machinery built around us. The harness decides what the model actually gets to see. It gives it the tools to act with. It runs a loop turn after turn. It keeps the history and applies whatever permissions you set. And when something breaks, it's the harness that has to recover and then check the result in the end. A while back there was a really interesting paper that might challenge your assumptions about what should go into a coding harness. And what they did is start with the bare shell, the model in a terminal with basically nothing wrapped around it. This setup scored 11 on their benchmark. Then they added their full harness, all the proper tooling and the exact same model jumped to 18 in their benchmark. Then they did something interesting. They took that full working harness and started changing one piece at a time just to measure what each design decision was actually worth. When they pulled out the editing tool, one single component, that score collapses down from 18 down to 10.3. That wiped out pretty much all of its gains and then some. And then the search was even sneakier. Swap that well-designed search for a clunky version that pages through results one at a time and you lose six points, which is actually worse than having no search tool at all. So, giving the model more isn't automatically better. Showing it the whole file instead of a focused 100-line window might actually hurt things rather than improve it. So, with these batteries-included agents, the likes of Cloud Code and Codex, you can see the sheer amount of stuff that these ship with straight out of the box. The long system prompt, the big bundle of tools, MCV support, sub-agents, permission pop-ups, plan mode, and lots more. And the list keeps on growing.

[00:03:24]There's a reason these are so popular.

[00:03:26]You just install it and it kind of just works out of the box. There's nothing to assemble and updates are shipped automatically. For a lot of people, that's exactly what they want. But, all of that machinery comes at a cost. Every feature adds additional bulk that you can't necessarily switch off. With these tools, you end up with a black box. And if something goes wrong, good luck trying to work it out. On the other side of the spectrum, you have Py. And when you use it, the first thing you'll notice is that it has access to a pretty limited set of tools. By default, pie gives you read, grep, find, and list, and then bash, edit, and write. It's a pretty limited subset of tools, but you can get a surprising amount done just with those alone. If you want to give pie instructions, then it can automatically read from an agent's that I'm D file, and your sessions within pie get stored as plain JSONL files, essentially a readable log of everything that happened, and you can branch off from any point, so nothing is hidden away. Then the provider config is the part that matters for a lot of people.

[00:04:22]You can point pie at any open AI compatible endpoint, which means a local model running on your own machine slots in just as easily as an API from a major AI provider. On a previous video on our channel, you can see Daniel running the pie coded agent along with the onslaught Quen 3.6 model running locally. But, what's not here is really the main point. There's no MCP out of the box, there's no sub-agents, no permission prompts, no plan mode, and no to-dos.

[00:04:49]The point is that you deliberately add these back in as you need them, either by getting pie to build these extensions itself, or you can look at their package catalog, where you can install extensions, skills, prompt templates, themes, and they're very easy to install. I'll be showing you that later.

[00:05:05]And what's not in pie out of the box is security or safeguards. The creator of pie didn't want you to be locked down to any one particular type of safeguard or permission prompts or anything like that. It's up to you to decide what to include. So, if you're running pie on your machine, you really need to put it into a sandbox and isolate it from your operating system, because out of the box, without adding any security or safety extensions, there's nothing stopping this from running very destructive commands on your machine.

[00:05:35]Pie is very easy to install. Go to their docs, start here, overview, quick start.

[00:05:40]Then you can install it on your machine using this NPM command. But, as mentioned previously, I'd highly recommend that you containerize it, such as using Docker. So, once you're using Pi, you can type {forward slash} login, and then choose an account like Codex, for example, and or else sign in with an API key. You're generally going to get much better value from the likes of a Codex subscription as it currently stands. I separately got Claude Code to set this up in a Docker container, so that Pi can only access anything within this workspace on my machine. Okay, now we can send the agent a message. Let's just say hello to start with, and now we get a response. And to be a bit cliché here, I just cloned a to-do app from GitHub into this workspace, and I'm just going to ask it to please summarize the repo. Summarize the to-do app repo.

[00:06:22]Okay, it's given us a response here. And while it did that, it was using its built-in tools such as find to read and to list the directory. But let's say your workflow needs a lot more than that, you can go to packages on their site, and you can sort by most downloads here. You can see that there are over 5,000 available packages here. We have Pi web access, MCP, context mode, sub agents. We even have Claude Code style dynamic workflows, and much, much more.

[00:06:50]So, let's say we want to add in web access and sub agents. We have a command we can use here. For this we can type Pi install and whatever the package name is. So, for Pi web access, I'm just going to copy out this command. I'm going to run that and let it install.

[00:07:04]Okay, it's installing web access, and we can also install sub agents. We have PI install, NPM, Pi sub agents, and I started a new session here, and now we can see the extensions of sub agents and web access are now showing up. So, now I can say summarize the to-do app folder use sub agents. Now you can see it has access to sub agents and can now dispatch them. So, the community has rebuilt many of the features that we know from Claude Code and Codex, and this is one of the real differences between the two philosophies, that we can add those as modular pieces on top of the bare core. But it's not all smooth. These extensions are built by different people to different tastes, and they can step on each other and conflict. And this brings me nicely to what Pyrus Pi is actually built for, because you don't have to live off that catalog at all. An extension is just a TypeScript file sitting in a folder, and they can do quite a lot. You can register a new tool, gate any tool call, reskin the UI, plug into a model provider, inject additional context, and add separate commands. You can simply just get Pi to write its own extensions, and that's how many serious Pi users actually run them. You start minimal, and then just get Pi to add an extension upon itself as needed. So, let's answer the most practical question. What coding harness should you really use? And for the majority of people, especially users that are not as technical, using coding agents like Claude Code and Codex are still the best option. They're shipping updates continuously, and for regular users, that's a good thing. You're keeping on top of coding edge features, and you can leverage what the builders of those harnesses are giving you. Where the likes of Pi starts to make sense is for a narrower group, power users, dev teams, enterprise setups, people who need serious control and to understand sandboxing of their environments, or people running local models, where the likes of Claude Code and Codex's system prompts are just far too bloated for them to be able to handle. And very importantly, you have to be willing to own the security and the containment.

[00:09:00]So, you need to really be isolating those environments and understanding what your agent really has access to.

[00:09:05]Now, of course, that is the case for all coding agents. There are horror stories of Codex and Claude Code deleting folders that it shouldn't, but the risk is far lower with them because there are a lot of safeguards in place already.

[00:09:17]With bare minimal coding harnesses, there really is nothing to stop it from doing destructive operations unless you actually build that into the harness.

[00:09:25]There's no universal best coding agent for every single model, so you need to just test it based on your own work. And if you want to dig deep into building coding harnesses specific to your own use case, then Pi has made that a lot more accessible. If you want to learn how to build expert-level AI systems from the ground up, then check out the link in the description to our course, The AI Architects. We've got over 50 lessons covering everything from agentic retrieval to harness engineering. You can launch faster with starter apps, get access to all of our courses and content, and join a global network of AI builders. Thanks for watching, and see you in the next one.