The company meeting: what the human never gives up

What I wanted to fix

In the previous post, I showed how I turned on Hermes Cloud to keep agents working in the cloud 24 hours a day with my laptop turned off.

Having assistants active both on my local computer and in the cloud solved my scheduling problems, but it quickly created another challenge: the lack of clear boundaries. When you start setting up agents and see how effortlessly they answer questions on any subject, your first impulse is to hand over every chore to them. It is tempting to think a group of digital assistants can run entire projects without regular supervision.

That illusion falls apart fast in practice. I work during the day, and at night and on Sundays I build with AI agents. My free time in front of the screen is scarce. Spending those precious hours fixing clumsy mistakes made by an unguided agent would be throwing energy away. Worse still would be handing automated programs access to personal accounts, credentials, or core technical decisions.

I needed clear house rules for my setup. On September 24, 2026, I paused my experiments and held what I called our “company meeting”: a quiet moment to sit down alone, align our guidelines, and define what belongs to the agents and what the human must never give up.

Quick summary

• In the previous post, I showed how I turned on Hermes Cloud to handle automated routines while my laptop stayed closed.
• On September 24, 2026, I established strict operational boundaries between myself and my AI assistants.
• Website code is written in Google Antigravity with me at the controls; agents do not operate Google tools on my behalf.
• For image generation or deep web research, an agent prepares a ready-to-use prompt and I run it manually.
• On that same date, I decided to build my personal website as a public journal to teach beginners what I am learning for free.

First: what an AI code editor is and what the meeting meant

For anyone unfamiliar with programming, the term “code editor” can sound intimidating, but the concept is simple. It is just a text editor designed to write the instructions that make up software, apps, and web pages.

Google Antigravity is one such editor, built with integrated artificial intelligence models. Instead of typing every single command from scratch, you describe what you need in plain language, and the system suggests blocks of functional code. You can picture it like a document editor with a capable writing assistant: you offer an idea, and it helps draft the paragraph.

The comparison stops there. In ordinary writing, a missing comma or a slight typo is easily forgiven by the reader. In computer code, a missing semicolon or a misplaced bracket breaks the entire page. Because of that, while artificial intelligence can generate code quickly, the person sitting in the chair needs to understand the logic and review every file.

The “company meeting,” on the other hand, was my playful name for a personal governance check. There were no colleagues or employees in the room. There was only me, the assistants I had built, and a clean sheet of paper to write down who was allowed to touch what.

What I used

On September 24, 2026, my working setup relied on three core components:

• A Gemini Pro subscription through Google Antigravity, operated solely by me to build pages and manage code.
• The assistants already configured in Grok Bot and Hermes to handle routine support and daily checks.
• A written guide outlining the operational boundaries and rules for each assistant.

How I split the work

During that alignment on September 24, I organized our workflow around four practical decisions.

First, personal accounts and logins stay strictly with the human. The agents do not operate Google tools for me. They have no access to passwords, master keys, or browser sessions tied to my identity. Any interaction with external services or third-party platforms requires my direct involvement. This prevents unintended automatic actions and keeps private data secure.

Second, website code is always developed with the human in command. All coding takes place inside Google Antigravity under my immediate supervision. I rely on AI models to suggest improvements and structure logical blocks, but I am the one reviewing each line, running local checks, and saving modifications. No agent modifies core project files in the background on its own.

Third, ready-to-use prompts for images and deep research. When an assistant needs an illustration for an article or an extensive research sweep across the web, it does not act on its own. Instead of opening browsers or attempting to generate images directly, the agent drafts a ready-to-use prompt (the exact query or visual description) and hands it to me. I review the text, make adjustments if needed, open the appropriate tool, and generate the asset myself. The agent suggests the path, but I press the button.

Fourth, the decision to build the personal website. Working through these boundaries made me realize how valuable it would be to document the journey. I decided to create my personal website, rodrigoesporto.com, as a public journal. The purpose was to share what I am learning, without hiding mistakes, technical limits, or costs, offering clean and free material for beginners who want to use artificial intelligence in daily life.

What went wrong and what I learned

In my early experiments, excitement led me to believe that AI agents could manage complete processes independently, even deciding which external tools to call from minute to minute.

That expectation failed early on. When you grant excessive autonomy to an AI assistant on open-ended tasks, it frequently takes circular routes, suggests unnecessary steps, or invents invalid commands. Furthermore, when you stop inspecting each file that enters a project, you lose grasp of the underlying system. If something breaks the next day, you have no idea how to repair it because you were not involved in building it.

The primary takeaway from this meeting was acknowledging the proper role of each side. Artificial intelligence models are fantastic tools for outlining ideas, organizing information, and proposing technical solutions. Even so, critical judgment, account security, and final sign-off remain human duties. Delegating a technical task is very different from giving up ownership of a project.

Costs and limits

Money-wise, the cost here was the Gemini Pro subscription through Google Antigravity, which I used as my main coding environment.

The practical limits of this approach were clear:

• The agents cannot see my local files outside their designated directories, and they cannot take actions inside my Google accounts.
• Output depends on my manual availability: because assistants deliver ready-to-use prompts for images and research, I have to set aside time at night or on Sundays to generate assets and review code.
• This dynamic is slower than unchecked automation, but it provides stability and eliminates unpleasant surprises.

What I would do differently

If I were starting over today, I would establish this division of roles before configuring a single agent on a remote server.

I lost days testing complex integrations and trying to make assistants handle chores that I could finish myself in two minutes through a browser. Writing down house rules early on saves mental fatigue and prevents an ecosystem from expanding into a disorganized mess.

Who this is for and who it is not for

Setting firm operational boundaries between yourself and AI agents is useful for:

• People who work during the day and spend their evenings and Sundays building projects with artificial intelligence.
• Anyone who wants to use AI for productivity gains without handing over credentials, personal data, or technical control.
• Beginners who want to genuinely understand how their tools work, rather than relying on automated black boxes.

This approach is not meant for:

• Anyone searching for push-button automation promises where software supposedly does everything without human effort.
• Projects attempting to outsource central decision-making to automated assistants.
• People who have no interest in testing, reading, and learning the fundamentals of their tools.

Wrapping up

The company meeting on September 24, 2026, brought needed clarity to my daily routine. Establishing that code is written in Antigravity under my supervision, that agents do not operate Google accounts for me, and that they hand off ready-to-use prompts created a stable and secure workflow. That solid foundation gave me the confidence to start building my personal website.

In the next post, I will share how I cleaned house: fewer agents, more clarity.

If you are figuring out healthy working boundaries with your own AI assistants, come say hello on X: @RodrigoesPorto.

Sources and links

• Official Google Antigravity and Gemini documentation: https://ai.google.dev
• My official profile on X: https://x.com/RodrigoesPorto