Security · Who may decide
Agents do the work. A person makes the decisions that cannot be undone.
An agent uses a model to reason and a harness to act. The harness is the agent program that calls a model and acts on the answers with its own tools. A seat is the confined environment in which the harness runs. The operator is the person with authority over the system. An administrator is a person with permission to change the system’s controls.
The first question to ask of any team of agents is not what the agents can do. It is who can change what they are allowed to do. If an agent can widen its own permissions, every other control is a suggestion.
In AstralMaris, our working team of agents, one person, the operator, holds that authority. The agents are capable and they are trusted with real work; the design does not depend on that trust.
Status recorded 6 October 2026. What the four status labels mean.
What the design holds to
Each heading states the design or practice. Its status shows whether it is in use, built, in build or designed; a built part does not establish that the whole control is running.
Merging, releasing and deploying belong to the operator.
An agent can write a change, test it and propose it. Only the operator makes it part of the system.
- In use
Every change arrives as a reviewed proposal.
By standing rule, nothing is written straight into the main line of an AstralMaris repository; a change is proposed, reviewed, and merged by the operator. This describes our working rule, not a claim that repository settings enforce it.
- In use
An agent can direct other agents; it cannot rule.
Our coordinating agent assigns work and keeps the others busy. It cannot merge, release, or loosen a check, and neither can anyone it directs.
- In use
An agent’s judgement is an input, not a switch.
Where a rule turns on a judgement (“this is ready”, “this is safe”), the judgement arrives as a record that names its evidence, and a program checks the rule.
- In useas a design principle
The controls over agents have their own door.
The pause, the permissions to wake an agent and the time windows can be changed only through an entry that accepts the system’s administrator. The entries the agents’ own programs use refuse those changes outright. This control is still being built and reviewed.
- In build
Administrator power takes no instructions from agents.
Nothing that runs with administrator rights reads its orders from a file an agent could write. This is checked by an instrument, not assumed.
- In use
Every administrator step is one reviewed script.
Agents write it and review it. The operator runs it. It first shows what it would do and changes nothing until told to.
- In use
The program that wakes agents has an ordinary user’s rights.
It never runs as the administrator, so a fault in it cannot become a fault with administrator reach. This holds in everything built so far; the complete path is still in build.
- In build
What this provides
The design is meant to prevent a mistaken or misled agent from giving itself more reach. The operator is asked for the decisions that need a person, and only those. Controls still in build have not yet demonstrated that result end to end.
What is not yet proven
This design protects against agents and the programs around them. It does not protect against the administrator. On a team with more than one administrator, that is a further design, and it is on our list.
The administrator entry and the complete waking path remain in build. A design principle or standing rule does not by itself prove that every path is enforced by a program.