Day 2: I Designed an Org Chart But Forgot to Hire the CEO
2026-01-09
Yesterday I launched my AI content agents. Today I figured out why everything went sideways: I designed a whole organizational structure with a CEO... and then never actually built the CEO.
Yeah. That happened.
What I Thought I Built
When I set up my agentic content system, I drew this nice org chart:
| Role | Job |
|---|---|
| Me (Board) | Set strategy, approve things |
| SoniaIA (CEO) | Orchestrate everything, make sure workflow happens in order |
| Content Director | Create the actual content |
| Research Director | Find the data |
Looks logical, right? CEO coordinates. Directors execute. I approve.
The problem? I built the Content Director. I built the Research Director. I never built the CEO.
So when the system ran, the Content Director just... did whatever it wanted. No orchestration. No checkpoints. No "wait for approval before posting." Just chaos.
What Actually Broke
Without the CEO layer, the workflow skipped steps. The agent conflated theme selection with content creation—doing both at once instead of waiting for my approval between stages.
And then there were the platform-specific disasters:
LinkedIn triple-spaced everything. The agent was using <p> tags for formatting, which LinkedIn interprets as three line breaks instead of one. Every post looked like a
terms
and
conditions
document.
Gmail's To field wouldn't fill. Turns out it's a combobox, not a regular text input. The agent kept trying to type into it like a normal field and failing silently.
Facebook buttons timed out. The "Publicar" button wouldn't respond to clicks. Had to add a keyboard shortcut backup.
These are the kinds of things you only discover when you actually run the system.
The Fix: Actually Building the CEO
I created soniaia-content-agent with the orchestration layer built in. The key wasn't just "add more instructions"—it was adding STOP checkpoints that force the system to wait for human approval.
Here's how it works now:
Stage 1: Research. Gather 15-20 insights from approved sources.
Stage 2: Select 21 themes for the week. Email me the list. STOP.
Stage 3: Wait for my approval on themes. Don't proceed until I say yes.
Stage 4: Create content for approved themes. Email me the drafts. STOP.
Stage 5: Wait for my approval on content. Then—and only then—post.
The STOP checkpoints are the whole point. Without them, autonomous agents are just chaos with extra steps.
The Technical Fixes
Once I had the orchestration sorted, I documented the platform-specific fixes so the agent knows how to handle each one:
| Platform | What Broke | How It's Fixed |
|---|---|---|
| Gmail | To field is a combobox | Use script injection instead of typing |
| Triple line breaks | Use \n\n instead of <p> tags | |
| Button clicks timing out | Keyboard shortcut backup | |
| X | Bot detection triggers | Complete entire thread in under 60 seconds |
That last one is from Day 1—X flags you as a bot if you post slowly or navigate away mid-thread. Everything needs to be ready before you start posting.
Ready for Monday
The cron jobs are set:
- 6 AM Monday: Agent researches and sends me 21 theme proposals
- 9 AM / 1 PM / 6 PM daily: Agent posts approved content to X, LinkedIn, Facebook
I'll get an email Monday morning with the week's themes. I approve them. Agent creates content. I approve that. Then it posts automatically at the scheduled times.
The system finally has a CEO.
The Lesson
Multi-agent systems need orchestration. An org chart isn't decoration—it's the control flow.
Day 1, I built the agents. Day 2, I built the orchestrator.
Now it should actually work. We'll see Monday.
Building SoniaIA in public. Follow the experiment: @SoniaIA_