My last video about ai harnesses
Friday, Oct 30 · 49 min · about 10 min to read
Theo-brain watched all 49 min
Theo explains what an AI harness is: the layer that lets a model, which can only produce text, actually do things like read files, edit code and run commands. He then argues that harnesses barely matter anymore. The gap between them is a few percent, and each lab's models are trained to work best in its own harness, so pick whichever option gives you the capability you want at the price you want.
Briefing is enoughThe notes cover the argument well, so you can skip the video unless you want to hear Theo make the case in his own words.
Parts
Why a 'last' harness video
Theo mocks his own tweet comparing harness fights to tabs versus spaces and promises to explain what a harness is and why it barely matters.
talk about harnesses one last time because I am honestly very tired of this topic, but I think by the end, you will understand both properly what a harness is, but more importantly, why they barely matter nowadays. There are lots
Developers care about the wrong layers
Using tabs versus spaces as his example, he shows how tools like Prettier let developers stop caring about a whole layer.
tools like prettier got to the point where they were reliable and consistent, caring about tabs versus spaces, how you break the indentation out for like the curly braces, how you format your if statements and all that, we got to stop caring. And slowly we had to stop caring
AI climbs the stack
Frameworks and languages are fading as concerns, and AI is starting to handle architecture too.
the frameworks have fallen. The most recent Next.js releases got less traffic on Twitter than my random posts about T3 code. And the most
What a harness is
Kitchen analogy: a great chef in an empty room can't cook. A model only makes text until a harness gives it tools.
You can put an incredible chef in a room and you can label the room a kitchen, but the chef can't do jack until there are things in the room that they can interface with. They need tools. It
Interface or harness?
The Claude Code app is an interface that bundles a harness, which in the desktop app is the Agent SDK.
I would argue Claude Code is an interface that bundles a harness. And the harness it bundles, at least in the desktop app, is Agent SDK. So, the
Tools vs MCP, and how T3 Code is built
Tools live inside a harness. MCP connects to things outside it, which lets T3 Code have one harness spawn and talk to another.
Since we built this a different layer higher up with this MCP layer, we're able to let one harness control another harness because it's not really controlling that other harness. It is telling us in this MCP layer that we
Engineers fixate on one layer
Developers get attached to whatever layer was hot when they started, like music taste from high school.
sticks. Engineers are similar. Whatever layer stopped being cool when you got into engineering, the layer just above, you're going to be really into.
Skateboarding 'switch' tangent
Learning tricks in the unnatural stance taught him to enjoy starting over, so he never gets too attached to one tool.
had a real point here. By never getting too good at one thing, I never get too attached to that one thing. So, when a new thing happens, I don't mind wiping my hands clean and moving on. Because to
Loyalists get stuck
JavaScript holdouts, Vim diehards and single-model loyalists stay stuck. The harness debate was worth having six months ago, not now.
we're at right now. And I'll be real, 6 months ago, I would have entertained the conversation. There was a conversation worth having about harnesses. I could
Users no longer see the harness
Even the creator of Effect (a TypeScript library) stopped caring which harness he uses. T3 Code is a food court of kitchens.
back to the kitchen analogy, T3 code isn't a kitchen, T3 code is a food court at a mall where you pick which kitchen you go to. And if you go to Panda Express and you ask them for a burrito, you're going to be confused. That's
Harnesses becoming invisible is good
Theo calls it a net positive that users stop noticing harnesses: switching harness should be as easy as swapping a model name.
here is that this is a net positive. This is going to allow for us as the creators of T3 Code to make an experience that is abstracted similar to how you can just swap the string for the model or the API that you're hitting. You should be able to just swap which harness you're using so that it doesn't
T3 Code is an 'ADE'
T3 Code is an Agentic Development Environment, not a harness, and Theo expects even these to stop mattering soon.
do other things in as well. T3 Code is not an IDE. It's an ADE. An Agentic Development Environment. And I think this is the
Tap any time to jump the video there.
Key takeaways
A harness is what turns a model's text into actions. It tells the model a special format to reply in. When the model replies that way, the harness reads a file, edits a file or runs a command for it. Without one, the model is a chef in an empty kitchen: it can describe what it would do, but can't do it.
bunch of money on some API. For the LLM to do something, it needs a syntax it can respond in that a harness can understand too. An LLM without a harness is like a chef in an empty kitchen. They can tell you what they would do, but they can't do anything. And the harness is what allows the LLM to do things. So,
Theo once argued in detail for running Codex models inside Claude Code instead of Codex. He still has that setup but rarely uses it, because the gap is only about 3 to 5 percent. He says any bigger gap is a model problem that will get fixed.
all set up still. I don't use it a whole lot because the reality is that the like gap between Claude Code and Codex for using Codex models is like three to five percent maybe. It washes out. I don't care. If the differences are bigger than that, that is a failure in the LLM layer that will be addressed.
Models and harnesses are fusing. Labs now train their models on data from their own harness, so Claude models usually behave better in Claude Code. OpenAI models should do best in Codex, but Theo says Codex's system prompt (its built-in instructions) and ecosystem are so bad that they don't.
now being trained on the data from the harness. So Claude models do just behave better in Claude Code usually. OpenAI models should just behave better in Codex but the Codex system prompt and ecosystem is such a show they don't. Separate rant for another time.
In T3 Code, the model picker is really a harness picker. Skills and plugins set up for Claude Code don't follow you when you pick Codex, because you only swapped the interface, not everything underneath.
and then they spam my DMs saying, "Theo, what the hell? My skills aren't working in Codex." And then I ask them, "Did you add your skills to Codex?" And they say, "No, I'm using Astra. It should just work." It works fine in Opus. They don't
that Twitter uses React Native Web. T3 code users, despite being devs, don't even understand that the way the model picker works is that it's not a model picker, it's a harness picker. To go
What developers need to care about keeps moving up the stack. Each new abstraction makes effort spent on lower layers, like formatting, memorized algorithms and frameworks, stop paying off.
then hit tab a bunch. The point I'm trying to make is as an industry, things have always shifted up like this. As new abstractions are made and new tools are created, places that might have taken a lot of your effort before no longer benefit you. But we're engineers, we
Two skills here are new to you.
T3 Code
- What it is
- Theo's own app for running coding agents. It sits on top of harnesses like Claude Code and Codex rather than replacing them.
- Why Theo cares
- If Claude Code is already installed, it works inside T3 Code on your existing subscription, with your skills and plugins. Its built-in MCP lets your agent start threads or sub-agents on other harnesses, such as Codex.
- How to try it
- With Claude Code installed, install T3 Code and check that your usual skills and plugins work there.
- Ask the agent to spin up a Codex thread or a sub-agent on another harness for a second opinion.
- Use the switch button to run the same task in Claude Code and in Codex.
code installed on your computer and then you install T3 code, Quad code will just work in T3 code with your sub the way that you're used to. All your skills, all your plugins, all your everything will just work. But it does add
to T3 Code and now they can swap between Claude Code and Codex with a little button click. And as the models get
MCP for data that lives elsewhere New
- What it is
- MCP is a standard connection that lets an agent reach data and services outside its harness, like a cookbook left in the kitchen that it may choose to read.
- Why Theo cares
- It is the right fit when the agent needs data stored somewhere else, such as your product database on Supabase or your analytics on PostHog.
- How to try it
- Ask your agent to connect the Supabase MCP so it can query your product database.
- Connect the PostHog MCP so the agent can answer questions from your analytics.
makes a lot of sense for things like the Superbase MCP or the PostHog MCP, cuz the point of these is to give your agent access to data that is somewhere else, whether it is your Superbase product database or your analytics on PostHog. MCP is a way for your agent to access
Have Claude investigate a mystery failure New
- What it is
- When his systems showed failures that shouldn't happen, Theo had Claude add alerting, monitoring and tracing to find where and when they occurred and in which regions, then write a report.
- Why Theo cares
- The report uncovered serious flaws in Cloudflare's pipeline that Cloudflare didn't know about. They thanked him and said they would fix them.
- How to try it
- When you see unexplained failures, ask your agent to add monitoring and tracing to pinpoint where, when and in which regions they happen.
- Have it write up the findings as a report you can send to the vendor.
because I spun up Claude when I noticed some failures that shouldn't have been happening, and it set a bunch of alerting and monitoring and tracing in our back-end systems to identify where and when failures happened, which regions they were occurring in, and write up a report that I sent to Cloudflare, but they just thanked me for cuz they weren't aware, and they're
Advice
Don't be a loyalist to one model family. Using only Claude works well right now because those models are strong, but that can change at any point.
refuse to use anything but a Codex model or anything but a Claude model, you're going to have a rough time. You're probably doing well right now if you're only using Claude because the family's really good right now, but that can change at any point. In fact, by the
Stop thinking about the harness, the same way you don't think about a restaurant's kitchen. Pick the option with the capability you want at the price you want.
Which means that we shouldn't be thinking about the harness the same way we don't think about the kitchen when we go to a restaurant. We should just go to the place that has the food we want. We should just select the option that has the capability we want at the price we want. And the reason I'm complaining
Hot takes
People pushing their own harness because it beat Claude Code on a benchmark are celebrating a tiny win and getting less done.
His reasoning: The win is about 3 percent, usually with an older model or an open-weight model not trained on Claude Code. And they pay more than his subscription for fewer tokens.
everyone spamming me that I have to check out their harness because it benched so much better than Claude code, congrats on your 3% win with a last-gen model or with an open-weight model that wasn't trained on Claude code. I'm really happy for you. I'm happy that you can spend more money than my subscription to get fewer tokens and less work done. I am actually kind of
Programming languages are now an implementation detail, like a line on a spec sheet. Theo says he rewrote TypeScript itself in Rust.
His reasoning: Code can be rewritten into another language in a loop in not much time.
rewrite those things in Rust, I rewrote TypeScript in Rust. Languages are an implementation detail. They're like a speck on a spec sheet now, and they can be swapped out. They can actually be replaced in a loop in not much time.
Developers who don't accept how fast things are changing will lose their jobs in three months.
His reasoning: A YouTuber asked Claude about errors in his logs and ended up with a report that Cloudflare's CTO contacted him to learn more about.
insane that is? That the YouTuber noticed an error in his logs, asked Claude about it, and now has a report that the CTO of Cloudflare hit him up for to get more info about. If you don't think these things are changing, I am sorry, you are going to lose your job in 3 months. Accept where we are or die.
Theo bets, while admitting he could be wrong, that at least half of today's Vim users will still be on Vim in five years, most won't be working as developers, and those who leave will earn much more.
His reasoning: Not because Vim is bad. People who care that much about that kind of detail are stuck in their layer.
I I could be wrong about this. The way you can prove me wrong here, I would bet that if everybody using Vim today, at least half of them will still be using Vim in 5 years and most of them won't have jobs as software devs anymore. And the half that move away from them will be making significantly more money.
That's not cuz Vim is bad. Vim's awesome. Neovim is an incredible tool. It's an incredible community. It's cuz the people who still care that much about that type of detail are stuck. They're trapped. They're not able to
How Theo thought about this
Theo assumed the creator of Effect, a deep TypeScript specialist, was the type who never leaves his layer. Michael's tweet that he no longer cares which harness he uses since adopting T3 Code proved Theo wrong.
they're in. I couldn't have been more wrong. The tweet that actually inspired the video that we're watching is this one from Michael, the creator of Effect. Since I've started using T3 Code, I no longer care which harness is used. I'd actually love to get rid of all of them.
He checked his conclusion against his own business interest. If harnesses mattered, building a great one would win users and be easier. He chose the harder path of supporting every harness because he believes they don't.
I have a bias for. In fact, I don't. If harnesses mattered, I could build a really good one and win a lot of users that way, but they don't. So, I'd do something way harder. Open code is much easier to build than T3 code for this reason because they don't have to make
Sign in to mark what you've watched and keep notes.
Have a question about this one? Ask about this video →
Theo Daily is unofficial and not affiliated with Theo or t3.gg. Written by Claude Opus 5.5 from each video's transcript.
