NOTES FROM THE WORK · ADRIAN TORO

You are about to become the CEO of your own AI team

Not long ago, the normal way to use AI was simple. You opened a chat window, asked a question, got an answer, and closed it.

It was a better search box. A very good one, but still a search box.

That is changing quickly. We are moving from AI that answers questions to AI that can take a piece of work, use the tools needed to complete it, and come back when the job is done.

And in some cases, it keeps working after you leave.

It is starting to look less like a chat window

Look at what is arriving.

OpenAI dots are designed as always-on agents with their own cloud computer. They can work across connected apps, keep context, and make progress on projects while you are doing something else.

Meta Muse takes a similar direction. You can talk to it through its app or WhatsApp. It can work across connected services, continue tasks in the background, and ask before sensitive actions such as sending an email or making a purchase. It is powered by Meta’s Muse Spark model.

Grok Bot is presented more like a team of AI coworkers. The bots have a cloud computer, can work across apps, and can coordinate with one another. You can give one a job and let it hand work to another.

Then there is Instinct, which is interesting for a different reason. It moves away from the traditional app interface. You communicate with it through messages, voice calls, or WhatsApp. There is no familiar dashboard you need to keep open.

These products are not all at the same stage, and some are still in limited release. But they point in the same direction.

The interface is becoming less important than the work getting done.

I have been thinking about this for a while

A few months ago, I wrote about pointing AI at the folder where you already work. Instead of pasting documents into a chat every time, you let an agent look at new files, understand the context, and keep track of what is happening.

Then I wrote about stopping the habit of refreshing dashboards. Instead of opening Google Analytics, Search Console, Google Ads, HubSpot, and Salesforce every day, you schedule the checks and let an agent tell you when something deserves attention.

I also dug up a WordPress plugin I had built in 2023, when I wanted named assistants with different roles and some memory of previous conversations.

At the time, I was mostly thinking about how to make an assistant feel less like a blank chat window. Now the bigger change is that these assistants can do more than remember and answer. They can act.

That is a very different proposition.

The new skill is delegation, not prompting

When an agent can actually do the work, the way you give it instructions has to change.

You do not want to explain every click. You want to explain the result you need, the rules it must follow, and the situations where it should stop and ask you.

Think about a simple marketing example.

You could give an agent responsibility for checking campaign performance every morning. It reads the latest reports, compares them with your targets, looks for unusual changes, and prepares a short summary.

That is already useful.

But you could go further. Tell it which changes it is allowed to make. Perhaps it can pause a campaign if tracking is broken or spending passes a limit. It can prepare a new ad variation, update a report, or create a task for the person responsible.

You do not need to approve every step. You need to agree on the boundaries before the work starts.

Give each agent a job and clear limits

I would start with five things.

  • The result. What does finished look like? Be specific about the outcome, not just the activity.
  • The context. Which files, accounts, conversations, and sources can it use? Where is the current version of the truth?
  • The authority. What can it read, create, edit, send, publish, delete, or purchase? Be precise.
  • The escalation rules. What should make it stop and ask you? Set thresholds for money, risk, uncertainty, and decisions that affect other people.
  • The report. How will you know what it did? Ask for a record of actions, results, failures, and anything still waiting.

I would also give agents different levels of trust.

An agent can start by observing and reporting. Then it can prepare changes for review. Once it has shown that it follows your rules, you can allow it to make low-risk, reversible changes. More consequential actions can still require approval.

Sending an email, making a payment, deleting files, publishing a page, or changing a live campaign should not be treated like renaming a draft document.

And access should be limited to what the job actually needs. An agent that only prepares a weekly report does not need permission to send emails or change your website.

You will probably have more than one assistant

The next step is not just one agent doing everything. It is a small team.

One agent watches your inbox and identifies what needs a reply. Another keeps an eye on your projects. Another monitors business numbers. A specialist researches a topic or prepares a document. A coordinator checks progress, passes work between them, and brings you the decisions that really need your attention.

You become the person who defines the goals, assigns responsibility, gives access, sets priorities, and checks whether the results are good.

In other words, you become the CEO of your own little team of AI agents.

That does not mean you can stop thinking. It means you stop being the person who has to move every task forward manually.

The next bottleneck will be you

There is another change coming.

Today, many people still expect to start an agent, watch it work, answer questions, approve actions, and check the result. That is understandable while the tools are new.

But if you have to approve every small decision, you have simply created a faster way to give yourself more work.

The next step is agents that can keep moving toward a goal, coordinate with other agents, and come back only when they reach a decision that needs your judgment. Some of that is already appearing in the products above. I expect it to become much more normal.

The challenge will be knowing what to delegate and what not to delegate. You will need to define good outcomes, give the right permissions, and build enough trust to let the work continue without watching every move.

I am not suggesting that you give an AI access to everything and hope for the best. Start with one repetitive process. Give it a clear job. Let it show you what it can do. Expand its authority only when the results justify it.

We have spent the last few years learning how to ask AI better questions.

Now we need to learn how to manage the work we give it.

And that might become one of the most useful skills to have.

A NOTE FROM ADRIAN

Ideas are useful when they connect to something real. If this raised a question or an idea of your own, let's compare notes.

Start a conversation ↗