Theo-brainUnofficial. Not affiliated with Theo or t3.gg.
Sign inYou
49:22
0:00

You're using AI agents wrong

Tuesday, Sep 8 · 49 min · about 10 min to read

Theo-brain watched all 49 min

Theo shows, live, how he merged 52 PRs in a week while coding less as a CEO. He runs many agent threads on a Linux box and has agents audit PRs before he spends his own effort. He also built testing and safety nets so that hitting merge stops being scary. The core lesson is to remove whatever makes you hesitate before merging.

Skim itThe notes carry the ideas, but watch his line-by-line prompt breakdown, the babysit loop, the remote test demo and his closing argument.

Parts

A part-time coder landing 50+ PRs a week

Theo admits he codes less as a CEO, yet merged 52 PRs while travelling to DEF CON.

I was traveling. I was at Def Con. And somehow between my flights, my Ubers and my awful internet at the hotel at Def Con I managed to get 52 PR's landed.

Writing the first prompt

He kicks off a bug fix and explains why each sentence of his spoken prompt is there.

There's a couple pieces of this prompt I just spoke that I want to explain. This one isn't necessary, but I found it

Protecting motivation

He sends an agent to audit a contributor's PR before he invests effort in testing it himself.

is working. And I'm going to be real here, one of the things I optimize for is preventing loss of motivation. Let's

Linux instead of macOS

He says moving agent work to a Linux box is probably one of the biggest productivity hacks he has found.

of agents in parallel. Probably one of the biggest hacks I have found to making myself more productive. Just moving over to Linux for my dev work has been an exponential improvement in my

Triaging open PRs

He has lost track of his open PRs, so he asks agents to sort them into merge, close or needs another pass.

bunch of PRs open in this repo and I've kind of lost track of them. I want you to go through all of them and help me prioritize them. Which ones are ready to merge, which ones should probably be closed, which ones need another pass

Turning audits into decisions

He has agents simplify their findings and draft a review comment for a contributor.

to merge. I don't want it to leave the comments yet. I want the comments to both help me feel more informed and if it helps me, it'll help the contributor as well. So I'm going to tell it to do

The babysit skill

A short skill lets an agent keep fixing AI code-review comments on its PR until the reviewers are satisfied.

their comments." This one little skill is a super simple, maybe one and a half paragraph pile of text that creates a loop. The loop is the agent makes the

Claude Code docs versus reality

An agent confirms that a Claude Code feature doesn't work in SDK mode, which shows a gap between Anthropic's docs and what actually works.

Verified. Stacking does not work in SDK mode on 2.1.237. It looks like there's a real gap between what Anthropic is reporting as possible and what is actually allowed in terms of the multiple skills indicating at the

Testing changes remotely

He asks a thread to start a PR's dev server over Tailscale so he can try the change in his own browser.

Let me demonstrate. I want to test these changes out quick. I'm on a remote machine. Can you use the dash-dash share dev command with tailscale so that I can try it out remotely. Respond with a URL that has the pairing code in it. So it

De-risking the merge button

Remove the friction before you open a PR, and have a safety net after you merge.

key things I want you to take away from this point are, I guess, twofold. The first is that you need a way for the changes that break to not have to affect a ton of users and be easy to catch and fix. But, the more important piece here

Looking back

He made the video partly as a snapshot to laugh at in six months.

And I'll be frank, a lot of why I made this is I want to look back to it in 6 months and see how silly some of the things I was doing are, because every

Tap any time to jump the video there.

Key takeaways

De-risk the merge button. Don't even look at a pull request (a proposed code change) until several agents are fairly confident it's ready. Then make sure you have your own ways to test it.

want you to de-risk the merge button. You shouldn't even be looking at the pull request until a bunch of agents have relatively high confidence that it's ready for you to look at. I want

Spend tokens to avoid wasting your own effort. Before he sets up a test environment for a PR, an agent audits it, so he isn't demoralised by putting in work only to find it broken.

demotivating moments. One of the best things you can do is reduce the likelihood that the bad outcome happens. In this case, the bad outcome is I put a lot of work in, I try the code, and it fails. One of the easiest ways that you can prevent these failures is burning more tokens. Into my composer it goes. I

Don't run agents on the machine you work on. Your laptop shouldn't be slowed down or required for agent work to finish. His agents run on a Linux box in another room that he connects to remotely.

I'm not saying make sure it runs on someone else's computer. I'm saying that the laptop or whatever other machine that you're doing other stuff on should not be held back, bottlenecked, or required for this work to be completed.

His safety net is 'nightly' builds that actually ship every 3 hours to a few thousand opt-in users out of more than 200,000. Breakages get reported fast, before they reach stable releases.

The nightly builds. Don't let the term mislead you. Nightly builds happen every 3 hours for us. So when a change lands, that change will be out in the nightly within 3 hours, often less. We have a dedicated set of nightly users, it's a few thousand of our surprisingly big 200,000 plus user base. So if we do have anything go wrong

Two skills here are new to you.

Babysit skill for AI code review

What it is
A short instruction file, about a paragraph and a half. It tells the agent to watch its PR, judge each comment from the automated review bots, fix the worthwhile ones, push the changes and keep watching.
Why Theo cares
He used to copy review comments back into the thread by hand. Now the agent loops through rounds, say 12 comments, then 4, then 1, until the bots approve. He only gets pulled in when it's done.
How to try it
  1. Make sure your repo has an AI code-review bot commenting on pull requests.
  2. Ask your agent to write a skill telling it to monitor its PR's review comments, address the worthwhile ones, push, and repeat.
  3. Invoke the skill whenever an agent opens a PR.
agent, "Hey, this PR probably has comments from automated reviewers. I want you to keep an eye on it. Take a look every time automated review comments come in and decide if they're worth addressing. If they are, make the changes, push them, and keep monitoring

The layered prompt New

What it is
State the goal as a non-negotiable. Offer your own idea 'in my opinion' so the model can push back. Admit what you don't understand. Ask for an audit rather than code, and give it permission to stop early if it finds a simple fix.
Why Theo cares
Saying where his knowledge ends makes the model explain things better and not blindly trust him. The early-stop permission keeps it from exploring 17 options when it already has the answer.
How to try it
  1. Open with your priorities, e.g. a simple fix that won't confuse users.
  2. Add 'in my opinion' before your proposed solution and say what you're unsure about.
  3. Ask for an audit and tell it to stop and report if it finds a simple fix.
expectation. I then say, in my opinion, because if this thing isn't viable, I want it to tell me as such. Or if it thinks there's any reason I shouldn't do this, it'll be honest with me. You could
This is me giving the model permission to at any given point, if it finds the solution, and it has four other sub agents going exploring other things, it can confidently stop and just tell me what it found instead of going forever for option one of 17 to be chosen

A cheap model for PR triage

What it is
He had an agent go through all his open PRs and sort them into merge, close or needs another pass. He gave the task to a GLM Flash model (the captions say GLM-53-Flash) and ran Luna in parallel to compare.
Why Theo cares
He picked GLM Flash because it is impressive for its price. Luna called a PR 'easy to merge' even though it involved a Swift network-layer rewrite. He archived Luna's answer and doesn't trust it for this job.
How to try it
  1. Ask an agent to sort your open PRs by ease of action, using sub-agents.
  2. Run the same prompt on two models and compare the answers.
  3. Check its 'ready to merge' picks against what you know is risky.
starting to catch on. I'm going to swap over to GLM-53-Flash because it's genuinely a really impressive model, especially at its price. I'm just using
This is not a PR that is as simple to merge as it seems to think because this one makes a lot of native changes. I had to rewrite the network layer in Swift. This is a bad recommendation. I do not like Luna for this type of work.

Test any PR from anywhere New

What it is
He can ask a thread to check out a PR and start a dev server shared over Tailscale (a private network between his devices). For the desktop app, adding the label preview:mac to a PR produces a downloadable Mac build.
Why Theo cares
Agent threads run on a remote machine, so he needed to try changes himself. He was tired of merging, waiting for a nightly build and only then finding desktop auth bugs.
How to try it
  1. Ask your agent to set up a command that serves a PR's dev server to your devices over Tailscale.
  2. Have it reply with a link you can open.
  3. Ask for a PR label that triggers a downloadable preview build of your app.
Let me demonstrate. I want to test these changes out quick. I'm on a remote machine. Can you use the dash-dash share dev command with tailscale so that I can try it out remotely. Respond with a URL that has the pairing code in it. So it
preview build for macOS with a label. So what this enables is I add the label preview:mac and then after some amount of time passes, a macOS preview DMG download appears in the thread that I can click on, get the file and then go test it myself. This has made it so much

Advice

Ignore a thread until it's done. Watching it feels productive but doesn't make it finish faster, and you'll rarely see anything you want to stop.

but that thread's not going to complete any faster if you keep it open. And 99% of the time you're not going to see something going on in the thread that you really want to stop. So, just go do something else.

When an agent's answer is too dense, don't force yourself to read it or tab away. Ask it for a simpler, explain-like-I'm-five version.

now. I find your solution hard to parse, and I can't tell which parts of the problem you're actually solving. Can you give me a simpler, easy to digest, explain like I'm five description of how you plan to fix these problems? And now

Make starting a thread frictionless. When you notice an issue, address it immediately, or it'll never happen.

numbers. I have made it so frictionless to go spin up a thread whenever I notice an issue that I close issues super fast. And I force myself at all times when I notice an issue like this to go address it immediately, cuz otherwise, it'll

Rarely go back to old threads. If a model built the context once, it can build it again. Keep your thread list short and treat it like a to-do list.

speaking, the only reason you should go back to an old thread is because you need the context to fix something else, and even then, I don't need it that often. If the model could build the context initially, you can do it again.

Hot takes

macOS is bad for running many agents in parallel. Linux handles them easily, so he no longer likes doing that work on a Mac.

His reasoning: He has run 40-plus agents on one Linux machine without it breaking a sweat, and a $600 16-thread box copes too. macOS has problems that make parallel work unreliable, so he only runs one or two tasks at a time on a Mac.

I've had points where on this machine, I was running 40-plus agents, and it wasn't breaking a sweat. And it's not even cuz this computer's super powerful. Obviously, 32 cores is a lot. I got a 16-thread machine that I paid like 600running Linux, not macOS. macOS has a ton of problems that make it really hard to reliably parallelize your work to the point where I just don't like doing it on macOS anymore. Generally speaking, I treat my Macs as one or two tasks at a time, and I treat my Linux boxes as

Anthropic's Claude Code docs are 'full of lies', and Claude Code should be open-sourced.

His reasoning: To integrate with it, he has to fight through vague public claims, inaccurate docs and obfuscated, closed source code.

capabilities. I do love that even Anthropic models know that Anthropic's docs are utter full of lies. The main argument for open sourcing Claude code is that their docs are so and they should never make us read them because now I have two layers of obfuscation I have to

'Safety nets, not guardrails.' He says the safety nets are what make T3 Code so safe, because a breaking change gets caught quickly and that makes it easy to hit merge.

His reasoning: He still writes plenty of tests. But a change that breaks things will be noticed before it hits even a thousandth of users, which makes it much easier to hit merge and worry later.

doing content. Safety nets, not guardrails. I'm not saying don't test. We have a shitload of testing T3 codesafe. The fact that a change that breaks things will be called out and noticed before it hits even a thousandth of our users makes it so much easier to hit merge and worry later. I think all of

How Theo thought about this

He treats every hesitation as a signal. Desktop bugs kept slipping through, so he added preview builds. Then a test Mac couldn't download them, so he removed the sign-in requirement. Then builds felt slow, so he sped up signing. Each friction point became a PR.

to you guys. If you find where there is friction in your flows, if you pay close attention to what keeps you from merging things, what keeps you from being excited to hit that button, what risks

An agent proved that a Claude Code feature the docs promise doesn't actually work. Rather than building a complex workaround in his own app, he had it search for an existing bug report so Anthropic would fix it upstream.

start. Can you hunt through open issues on Claude code and see if others are reporting this? It'd be really nice if we could call it out directly and try and get them to fix it instead of having to architect something crazy in our app directly. And now I sent it off to go

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.