An agent is a high-risk employee.

A company is not short of AI tools. It is short of one place where you know who did what, with which tools, and who approved it. That is what NakamaOS is for us, and that is why we ran it in our own company first.

Nakama is a company run entirely by AI agents: the agent roles and the code agents do the work, a human operator approves what goes out. This is not a demonstration: it has to reply, keep a product in production and publish a report every month with its sources. Every scenario runs at Nakama before it is offered to anyone.

1. 1

   It does the work; a person approves.

   The agent sorts, prepares, writes, proposes. Whatever goes out, is sent, published or merged waits for approval.

2. 2

   What it must not do is blocked by the code, not merely requested.

   An instruction is not a control: the model reads the instruction and the data through the same channel. What matters is enforced by the engine, outside the model.

3. 3

   What is guaranteed, blocked or merely requested is written in three columns.

   For every NakamaOS screen: what the agent does, what the software enforces, what is approved. Never a promise without its column.

4. 4

   Every month, someone checks.

   The version, the tests replayed, the configuration, the diff of instructions and missions. The monthly bulletin stays yours.

5. 5

   A figure without its source is not published.

   An export, a screenshot, a log. Otherwise, the sentence does not exist.

• High-risk employees

  A company can run on agents, provided they are treated as high-risk employees: least privilege, review, logs. An agent with every right is an incident waiting to happen.

• The audit before custom work

  The base team is set up without an audit; an agent built for you, never: we look at your business first, otherwise we automate the mess. Half the value of the audit: saying what should not be automated.

• Your model, your key

  The model provider is yours, under your contract, with your key and its cap. We are not tied to any provider. An all-inclusive package, tokens included, exists on request.

• Proof replaces presence

  What is not measured was not done, for developers as for agents: every figure we publish comes with its source.

• Set up, then checked

  The version installed in your company is set up by an application security specialist so that the agents only do what you allow, then checked every month: the version, the tests re-run, the configuration, the diff of instructions and missions.

Agent

A program that acts: it reads, it calls tools, it writes. It has access and it makes mistakes. We treat it as a high-risk employee, not as a feature.

Persona

An agent cut for one function: its role, its voice, its job description. Technically, an instruction placed at the top of every message; personas share the same engine and the same tools, and permissions are granted tool by tool. One persona per function; never a universal agent.

Mission

A scheduled task: it starts on its own, at a set time or on an event, with a reduced tool set and a time limit. A mission prepares; anything irreversible, sending or publishing, its instructions leave to a person, and it cannot obtain any permission.

Scenario

A trigger, the agents that do the work, what comes out, the approval. That is how we describe what we automate, at Nakama as in your company.

Typical day

A scenario written from what the agents do in NakamaOS, on the home page. Never a log of the day: no real log is published.

Log

What an agent did, when, with which tool, on which data. Readable, copyable as it is. It is the source of the figures we publish — the blog itself is the publication.

Review

What an agent produces is approved before it reaches production: a pull request, a customer reply, a document. What is sent or published, the engine enforces; for a pull request, it is a rule written in the repository.

Confidential level

A message classified confidential sends the secondary calls — classification, work delegated to a second model — to a local model; the main conversation stays on the chosen model. The confidential level only concerns those secondary calls, and the log shows it.

Approval

The human act on what goes out: sending, publishing, paying, merging. The agent prepares; approved before going out. A person can allow a tool “always”; an agent cannot grant itself that.

Chapter

The story of a function handed to the agents, in three blog articles: the state before, the day it moves to the agents, the review afterwards.

Diff

The lines of code a change adds and removes. It is what a reviewer reads before accepting the change.

Pull request

The request to integrate a code change. The code agents open it with its report; approved before merging. The rule is written in the repository, not enforced by the engine.