How a synthetic officeof agents is governed

01What a synthetic office is

A synthetic office is a team of AI agents organized like a company: with a hierarchy, mandates by area, and someone who coordinates. It is not a program that does everything. Each agent has an area, knows what is theirs, and knows who to report to. Anthropic describes this way of working as the orchestrator-workers pattern: a central model splits tasks among others and gathers the results¹.

Every area agent works with a director's mandate: what falls to it in leadership, customers, resources, and the other dimensions of the model. The company gains the structure of a full leadership team, and the people who lead today keep deciding.

02The governance rules

There are six, and they are not negotiated project by project. Each one answers a concrete risk.

1. The final word belongs to a person. The general director and whoever leads each area decide. Agents execute, prepare, and propose.
2. Every task has exactly one owner. Every request falls into a cell of the mandate matrix, and that cell has one responsible agent, as in a RACI matrix, where every task has exactly one person accountable for the result³. If there is no cell, the request goes to a person.
3. Compliance does not depend on anyone it reviews. An auditor who reports to the area it audits stops being an auditor. The Institute of Internal Auditors' three lines model puts it this way: independence from management is what gives audit its objectivity and authority⁴. That is why Compliance watches every vertical and reports directly to the coordinator.
4. Read by default, write with permission. Every agent starts with read-only access to each system. It writes only where your company authorizes it in writing, system by system, and asks for authorization before each write. That is the principle of least privilege: everyone gets only the permissions their function needs⁵.
5. Whatever moves money or commits the company is approved by a person. OWASP recommends human approval before any high-impact action⁶. In our model that includes payments, signatures, contracts, and legal or tax decisions of substance.
6. Everything is logged. Every action records when, where, who, and what⁷. Since every agent has its own account, "who" is always identifiable.

03The decision-rights matrix

This table is the core of governance. It says what the agent does and what the person does for each kind of action.

The practical effect is that your team stops chasing information and starts receiving it already put together. Whoever approves a payment does it with the invoice validated, the purchase order cross-checked, and the supplier's history in view.

04How a request travels

A request always follows the same path, and every step leaves a record.

1. It comes in. By WhatsApp, email, a form, or a system, at any hour.
2. It gets classified. The CEO agent places it in a dimension and sends it to the agent who owns that cell.
3. It gets executed. The agent works with its permissions and prepares what needs a decision.
4. It gets reviewed. Compliance checks that what was done follows the rules before anything goes out.
5. It gets approved. If it touches money, a signature, or a legal criterion, a person decides.
6. It gets closed. The case is logged, and shared memory keeps what was learned.

Anthropic warns that an agent's errors can compound and recommends pauses for human review and stopping conditions¹. Steps 4 and 5 are those pauses.

05Infrastructure and data for each client

Every client has a dedicated virtual machine, Claude account and agents, and shares infrastructure with no one. That infrastructure belongs to Omni Agent, which contracts and manages it as part of the service; the information stays the client's.

Agents connect through authorized channels that already exist: banks through authorized aggregators in read mode, SAT through bulk invoice download, the ERP through its own connection or a connector installed on your network, and customers through WhatsApp Business and email. Where a portal has no connection, the agent operates it the way your staff would, with the permissions you give it. Your data stays yours. The confidentiality agreement is signed before seeing the first document, and personal data handling follows the federal law on the protection of personal data held by private parties, whose current version was published in March 2025⁸.

06How we install governance in your company

Omni Agent is an AI consultancy based in Mérida, Yucatán, that installs agents on top of models that already exist, mainly Anthropic's. Governance is not added at the end: it is designed from day one.

1. Discovery. We map who approves what today, the business rules, the policies nobody wrote down, the tools, and how many times a month each thing happens. You come out with your company's decision-rights matrix.
2. Contract. If you decide to move forward, it is signed once Discovery ends, for twelve months.
3. Implementation. We set up each agent's account, permissions by system, approval points, and Compliance controls, and we test with real cases before opening the flow.
4. Maintenance. We review the log, adjust rules and permissions, and give your team one hour of support each week. The monthly fee also covers the Claude account, the virtual machine and the harness license.

With governance installed, your directors delegate more because they see everything. The numbers are current, information arrives already put together, and decisions get made in minutes instead of waiting for the month's close.