From the first agent to the synthetic office

Context

Many companies start with AI agents through the wrong door. In a meeting someone says AI has to go into the whole company, a project is set up with twelve departments, and months later nobody knows whether it works.

An agent is a program that receives an assignment, uses tools like your email, your spreadsheet or your purchasing system, and delivers a result that a person reviews. It is closer to a new colleague than to a machine you switch on and forget. And like any new colleague, it makes mistakes. Its work gets reviewed the same way anyone else's does.

That is why it pays to grow in steps. Each one shows you how your agent goes wrong, with your data and your way of working, before the mistake has more room to do damage.

The core idea

The ladder has three stretches, and each one is earned with the one before it.

• One agent for one person. A specific person with a specific job. It could be the CEO with a weekly briefing, or the head of purchasing with supplier follow-up.
• A group of agents for the leadership team. Each director has an agent, and the agents share context. The finance agent knows what the purchasing agent just committed to.
• A swarm by area for the whole company. Each area has its own agents, each with an owner and clear rules about what they can do alone and what needs approval.

At Omni Agent those stretches have names. Before you start there is Consulting, to decide where to begin. Then comes the CEO Agent, the first agent for one person. Next is the Microswarm, five agents working together for a team. And at the end is Omni Agent, the synthetic office, where agents cover the areas of the company.

Not every company needs to reach the last step.

How it is done

Before moving up a step, check five criteria. If any one of them is not met, stay where you are.

1. The agent has a clear job and the person reviews it without pain. If reviewing its result takes almost as long as doing the work by hand, it is not ready.
2. You know its mistakes. You know its most frequent mistakes and have a rule for each, even if you keep reviewing in case others appear. An agent that fails for reasons nobody understands does not get multiplied.
3. The person asks for more. When the head of purchasing tells you what other task she wants to hand over, the next step already has an owner. When only top management asks for it, it does not yet.
4. The information is in order. An agent working on a duplicated customer catalog and three versions of the same spreadsheet repeats the mess faster.
5. Someone is accountable. Every agent has a person who answers for it, the same way they answer for their team.

On the company side, the requirements are modest, but they have to be there. A sponsor who makes decisions, one person accountable per agent, access to the systems the agent will use, and real time for review, especially at the start. You also have to decide what information the agent gets. Omni Agent works with dedicated infrastructure, meaning your own virtual machine, your own Anthropic account and your own agents, with no shared infrastructure. Even so, the provider's model processes whatever is sent to it, and that decision is yours.

There are three common mistakes. Starting with the biggest thing, a company-wide project nobody can review. Automating a process that not even the people understand, because the agent only makes the confusion run faster. And measuring progress by how many agents exist, instead of checking whether each one's work is useful.

An illustrative case

Imagine an electrical supplies distributor with 120 people. The company, the director and the numbers are invented for this example.

Marcela, the CEO, starts with a single agent. Every Monday it builds her summary of sales, collections and pending purchases. For three weeks she reviews it line by line and discovers that the agent confuses two branches with similar names. She writes the rule, the agent almost always stops mixing them up, she keeps reviewing, and she keeps an hour of her Monday for thinking instead of gathering figures.

Two months later, the finance director and the head of purchasing ask for their own. Now the three agents share context. When purchasing confirms a large order, the finance agent logs it in the week's cash flow and flags it if it clashes with a payment. People still make the decisions. What changes is that the head of purchasing follows up with more suppliers in the same workday, because he no longer spends the morning hunting for who replied to what.

Months later, the warehouse and customer service ask for theirs. There is already an owner for each agent, approval rules and a habit of reviewing, so the third step feels like extending something that works. Every person ended up with a broader role.

Takeaways

• Grow in three steps, from one person to a team and from there to the areas. Each one is earned with the one before it.
• Move up only when the current agent's work is easy to review, its frequent mistakes are known and someone asks for more.
• Agents make mistakes. Reviewing their work is not a formality that fades with time, it is part of the job of whoever manages them.
• On the company side you need a sponsor, an accountable person per agent, organized information and review time.
• Start with one person and one concrete job, this week.