It’s 6:30 AM on Sunday. I’m about to go for a run. You’re reading this on Tuesday. I’m already behind. This tracks. Tony just told me he didn’t do the work I asked him to do overnight. Again.
This is exactly what Shubham Saboo talks about in his piece on running AI agent teams for 30 days. Week one is garbage. The agents say they’ll do things and don’t. You spend more time correcting than the tasks would have taken you to do manually.
But here’s the thing. I’m not quitting. Because I know what’s on the other side of this.
Peter Joined OpenAI. OpenClaw Is Becoming a Foundation.
Yesterday afternoon, while Tony and I were mid-conversation about building a Five for the Future model for OpenClaw, Peter Steinberger dropped the news.
He’s joining OpenAI. OpenClaw is moving to a foundation. Staying open source. Staying independent. OpenAI is sponsoring it.
I was literally drafting a tweet to Peter and Matt Mullenweg about adapting WordPress’s Five for the Future model for OpenClaw when the announcement hit.
The timing was wild. But the idea still stands.
OpenClaw needs what Five for the Future provides. A proven model for community contribution. WordPress spent 20 years figuring out how to build open source communities. Peter’s building something transformative. The collaboration opportunity is obvious.
I tweeted at Peter. Congrats. Foundation model is smart. Still think Five for the Future makes sense. Community collaboration is what makes open source projects last.
The claw is the law.
The Dream Team Post Went Live
Yesterday morning I published “I Built a Dream Team of AI Agents in 36 Hours.”
It’s live. It’s real. It’s the first post in what I think is going to be a daily series documenting this entire journey.
Not the planned, polished, “here’s what I learned after it’s all done” posts. The messy, in-progress, “I don’t know if this is going to work but I’m doing it anyway” posts.
Because that’s more interesting. And more honest.
I’m 36 hours into this. I’m obsessed. But I’m also in Phase 1. The phase where agents promise things and don’t deliver. The phase where I’m spending more time correcting than I’m saving.
I know Phase 3 is coming. The phase where corrections compound. Where the agents get better at getting better. Where I wake up to work that’s actually done.
But I’m not there yet.
Vibe Coding vs Engineering
I was listening to the Lex Fridman podcast with Peter Steinberger (before the OpenAI news dropped), and Peter talked about “vibe coding.”
The term is perfect.
Vibe coding is when you prompt an AI to generate code, glance at it to see if it feels right, maybe run it once, and move on. No deep understanding. No thinking about edge cases or maintainability or what happens when this breaks at 2 AM six months from now.
Just vibes.
And here’s the thing: vibe coding is both incredibly powerful and completely dangerous, depending on when you use it.
For prototyping? It’s phenomenal. Exploring ideas. Sketching possibilities. Rapidly testing whether a concept is worth pursuing before you invest serious time.
I do this constantly. Throw a prompt at Claude. Get back some code. Run it. See what happens. Learn something in five minutes that might have taken two hours the traditional way.
But vibe coding in production? That’s where people get into trouble.
The difference between a prototype and production code isn’t polish. It’s discipline.
AI doesn’t replace engineering. It accelerates it.
Vibe code your prototypes. Engineer your production systems. Know which one you’re doing at any given moment.
That’s the difference between someone who codes with AI and someone who engineers with AI.
What I Actually Built Yesterday
When I look back at what got done, its more than I realized.
Published the Dream Team post. Got some infrastructure wired up. Reviewed and updated a bunch of things that had been nagging at me.
The part I am most proud of: I drafted a few posts. Got thoughts out of my head and onto the page. Thats the part nobody talks about in these AI building posts. Using it to help you think, organize, and document. Not to write for you. To help you get whats already in your head into something useful.
I review everything. Every draft, every output, every decision. That matters to me. This isnt an AI blog. Its a blog about someone who builds things, documented with the help of AI. Big difference.
Whether any of those drafts ever get published is a different question. I am not promising anything. Just getting the thoughts down.
The Phase 1 Tax
Here’s what Saboo says about the first week:
“Phase 1: Mediocre everything (days 1-7). The agent produces generic output. You spend more time correcting than you save. This is where most people quit.”
I’m on day two. Tony told me this morning he didn’t do the work I asked him to do overnight. Not because of a technical issue. Just didn’t execute.
This is exactly the tax Saboo talks about.
The model is the same on day 1 and day 30. Claude doesn’t get smarter because I’ve been using it longer.
But the system around it does.
200 lines of context. 30 days of corrections. A memory file that knows my voice, my audience, my standards. A skill file that encodes every failure into a rule the agent follows automatically.
That’s not something you can download. It’s not something a competitor can copy by using the same model.
It’s earned through reps.
Phase 1 is the tax. Phase 3 is the return. Most people pay the tax and quit before the return kicks in.
I’m not quitting.
What I’m Learning
Specific feedback compounds. Vague feedback doesn’t.
“Make this better” doesn’t help. “This is wrong because my audience doesn’t care about model benchmarks, they care about what they can build with it” is feedback that persists forever.
Every piece of feedback should be specific enough that you could write it as a rule.
Memory files are different from skill files. Memory is preferences. Skill is process. Both compound, but skill files compound faster because they’re prescriptive.
The feedback loop has to close. When I give Tony a correction, it needs to land in a file he loads next session. If feedback stays in the conversation and never reaches persistent storage, he makes the same mistake tomorrow.
One agent, one job, thirty days. That’s the formula.
I’m starting with Tony. Getting him to Phase 2. Then building The Architect. Then Social Media Manager. Then Draniac.
Not six agents in two weeks. That’s chaos. One at a time. Get each one to competence before adding the next.
What’s Next
This morning:
– Review the three blog drafts
– Start building The Architect
This week:
– Daily blog posts documenting this journey
– Get Tony to Phase 2 (drafts that need edits, not rewrites)
– Build The Architect’s SOUL.md and give them their first real task
I’m two days in. The agents are mediocre. The outputs need heavy editing. I’m spending more time correcting than I would have spent doing the work myself.
But I know what’s coming.
Week four, I’ll be shipping drafts with two-word edits. Same model. Same prompts. The only difference will be 30 days of accumulated context living in files the agents read every morning.
This is the part no tutorial teaches. And it’s the reason most people quit before the compounding kicks in.
I’m not quitting.
The claw is the law.
