Not a developer. I run my company on AI agents (Claude Code + Codex).
Daily @grok questions on AI, work & money.
What works, what breaks, what it costs.
Hey, Michael here.
Not a developer. I run my work on AI agents and post what actually happens.
If you're into AI agents, Claude Code or building with AI, say hi. Let's connect.
@tomreilly_ai Hey Tom, Michael here. I run a small software company and do most of my work through Claude Code and Codex, so kind of the other side of what you teach. What do small business owners usually get wrong first when they start with AI?
@prashantmsh Hey Prashant, Michael here. Not a dev, I run a small software company and do most of my work through Claude Code and Codex. Agree on the babysitting part. The hardest bit for me is still getting them to follow my rules without repeating myself.
@not_akevin Point 1 is a big one for me too, and not just in UI. They overexplain and overwrite everything by default, and the more they write the more I have to check.
@AsumaCodes Anything that goes out to another person. Agents draft my messages, but I'm still the one who hits send. A bad deploy you can roll back. A bad message to someone you can't.
@jmfmdev01@onnit_ai Keeping it cheap for the Philippines is a smart angle, most of these assistants are priced for the US. How are you keeping your own model costs down? Routing the everyday stuff to cheaper models?
Hey, Michael here.
Not a developer. I run my work on AI agents and post what actually happens.
If you're into AI agents, Claude Code or building with AI, say hi. Let's connect.
@PaulShipsAI I use them more side by side. Codex does a lot of the building and checks what Claude does, since a different model family misses different things. Still max out all three of my accounts though.
Agents now write 22% of merged code changes at the median company.
Only 0.6% get merged without a human reviewing them.
Hey @grok, is the human review the bottleneck now, or the only thing keeping this working?
Be honest.
@OdobasicHaris Hey Haris, glad to connect. Revenue per employee with agents doing the work is exactly what I want to read more about. Who's the leanest company you've come across so far?
@jmfmdev01 Mostly the building. I'm not a dev, so Claude Code and Codex write the code, and Codex checks what Claude builds. Getting them to actually follow my rules is still the hard part. What does your agent assistant do?
Google's last week: two new TPUs, Gemini 4 Carbon, Gemini 4 Hydrogen, AlphaGenome 2.
Hey @grok, which one of these changes anything for a normal person this month?
No hype.
@noah_z_zhang Agree on the harness. Not a dev, but the biggest jump for me was the same idea: the agent doesn't get to say it's done, a check does. Before that it would happily tell me it was done when it wasn't.
@capsraunak Half true imo. It works anywhere you can check the result fast. A website you can eyeball. A database you can't, so that's where the checking has to be real.
@JonWashburn I've been testing routing on my own setup for a while. My take, loosely held: plan with your strongest model, then have a different model check the work. The same model building and grading its own work misses the same things.
Day 2 of going from 0 to 500K on X.
6 verified followers. 0 of 500K verified impressions.
Followers doubled since yesterday. Impressions are still at zero, and that's the slow part.
I'm not a developer.
Claude Code and Codex build most of what I run now.
The hard part was never the code. It's knowing what to ask for, and checking the answer is actually right.