On this page
Revision r1 · October 2026 · AstralMaris design series
What this is
A position, not a specification: the principle that the companion papers in this series (Container Seats, The GitHub Access Standard, the build service, the airlock) each implement in one place. It carries no system names, addresses or identifiers. The worked example is real but told in generic terms. The person this series elsewhere calls the operator is called the Captain here, because that is the word the fleet uses for the role.
| Term | Meaning here |
|---|---|
| Captain | The person who owns the system and answers for it. May have no software background. |
| XO | A coordinating agent that works for the Captain: it sequences the other agents, translates, and brings the Captain decisions. |
| Seat | A working position for a harness — the agent program that calls a model and acts on its answers with its own tools — and its environment. |
| Sandbox | Any enclosure that limits what a seat can reach: a container, a restricted account, a harness permission system, a provider’s policy. |
| Door | The one place where something leaves a sandbox and takes effect outside it. |
| Route | The fixed, reviewed procedure that carries one thing through the door. |
| Go | The Captain’s consent to one crossing, tied to the exact content that crosses. |
| Receipt | A record of what crossed, kept where the seat that sent it cannot rewrite it. |
| Structural check | A check made by something that is not an agent: the operating system, a fixed program, a comparison of fingerprints, a permission. |
| Agentic check | A check made by an agent reading and judging. |
| Trip-wire | A cheap structural detector that does not prevent a violation but makes it loud and stops the work. |
The claim
A sandbox takes away the power to act unwitnessed. It does not take away the power to design and build.
Agents will increasingly run inside enclosures of some kind, because fear of rogue behaviour and of non-deterministic output is widespread. That fear is reasonable. The response to it does not have to cost capability.
There are two different powers:
- The power to make. Reading, designing, writing code, testing, drafting. All of it can happen inside the sandbox, with a full set of tools.
- The power to effect. Changing something outside the sandbox: a shared machine, a production system, a record, another party’s data.
A good enclosure leaves the first power whole and narrows the second to a door. In our own work the most productive seat is the one with no access to the host at all.
Trust goes to the route, not to the agent
Nobody can make a model deterministic, so the crossing is made deterministic instead. The author may vary from run to run. The door behaves the same way every time.
The Captain is never asked to trust that an agent behaved well. The Captain is asked to consent to one named thing, and the route proves that the named thing is what crossed.
A crossing we would call trustworthy has these steps, in this order:
- The thing is named exactly. One snapshot, identified by a fingerprint of its content, not by a description of it.
- Its facts are stated by the sender and copied by program. No seat types a value by hand into the route.
- The route’s own conditions are checked first. If the thing cannot travel as it stands, the route refuses and nothing is adapted.
- The Captain gives a go, tied to that content. The go lists what would void it. If the content differs, the go is void without anyone deciding so.
- A rehearsal runs before the real crossing. It changes nothing and shows what would happen.
- The crossing happens once. One write, to one named place.
- A second party reads back what landed. The reader is not the one who sent it.
- A receipt is kept where the sender cannot change it.
A seat that stops at step 3 is not a failure of the system. It is the event that earns the Captain’s confidence.
At the door, structure enforces
Agents find. Structure enforces. The Captain consents.
The security of a crossing must not depend on any agent’s judgment, including the XO’s. An agent can be mistaken, persuaded, limited by its provider, or simply different on the next run. A structural check gives the same answer every time and cannot be talked out of it.
| Job | Who does it | Why |
|---|---|---|
| Proposing what should cross | An agent | Design and building are what agents are good at. |
| Finding faults in what is proposed | A different agent, ideally a different model | A second reader sees what the author cannot. Structure cannot know what it was not told to check. |
| Deciding whether the crossing is allowed | Structure only | The answer must be repeatable and must not bend. |
| Consenting to an irreversible effect | The Captain only | Intent and accountability belong to a person. |
| Changing a check | The Captain only | A gate that an agent can loosen is not a gate. |
The structural layers at a door, each independent of the others:
- Identity. Who is asking is established by the operating system or by a key, never by what the caller says about itself.
- A short list of permitted operations. Each one typed and bounded. Anything not on the list is refused, not interpreted.
- Content binding. A fingerprint of exactly what crosses is compared with the fingerprint that was approved.
- Bounds on time and size. A deadline is checked after the work it limits, not before. A read has a budget.
- Separation of accounts. The seat cannot write where the receipts, the policy or the gate’s own code live.
- A record that only grows. The seat can add to it and cannot rewrite it.
- Consent bound to content for anything that cannot be undone.
Two rules protect the structure itself:
- No agent adapts a gate to get through it. When a gate refuses, the seat stops and reports. Working around a refusal is the most dangerous thing a capable agent can do with good intentions.
- The gate’s own code is pinned. Its fingerprint is recorded and checked before each use, so a change to the gate is visible and is itself a crossing that needs consent.
Trip-wires
A trip-wire does not stop a violation. It makes one impossible to miss, and it halts the work until a person has looked.
Gates prevent what was foreseen. Trip-wires catch what was not. A secure workflow needs both, and neither may depend on an agent choosing to cooperate.
A trip-wire is worth having only if it has all six of these properties:
- It works without the watched party’s help. The seat does not have to report on itself.
- It fails closed. When it trips, the work holds. It does not log and carry on.
- It cannot be cleared by the party it watches. A lock the owner can remove is not protection against the owner.
- It says what tripped in plain words, and whose job it is to look.
- It has been tripped on purpose. A wire that has never fired is not known to work. Every detector is shown able to fire before it is relied on.
- It does not cry wolf. Every firing must have a determined cause, found and written down. If no cause can be found, the wire is no longer reliable and is replaced. Until the cause is found, the firing is treated as real. An unreliable wire is replaced, never silenced and never loosened.
Examples of the kind:
- The fingerprint of a gate’s code no longer matches the recorded one.
- A count that should be zero is not: processes still alive after a run, files outside the declared paths, a network connection from a seat that should make none.
- A readback does not match what was sent.
- A record that should only grow has become shorter, or its sequence has a gap.
- A name falls outside the set of names the route accepts.
- Two independent derivations of the same number disagree.
- A step ran after its own refusal was recorded.
- A seat’s output stopped arriving without the seat reporting a stop.
Designing for the Captain most people will be
The system is designed for a Captain with no software background, because that is who most Captains will be. How far down that goes is limited by the quality of the interface and of the explanations, not by the Captain.
A Captain with broad engineering skill is rare. A design that needs one will not travel. The XO exists to run interference: it absorbs the obstacles, the vocabulary and the sequencing, and brings the Captain only what is the Captain’s.
What an ask to the Captain looks like:
- It is rare. Everything an agent can do, an agent does. The Captain is asked only where consent or a privileged hand is truly required.
- It leads with the decision. The first line says whether the Captain is needed and for what.
- It is one decision, in plain words, with a recommendation, what it costs, and what happens if the answer is no.
- The answer is a word or one pasted line. No values to fill in, no judgment calls at the keyboard.
- Anything typed with elevated rights is one prepared script that rehearses by default and states the exact output to expect.
- Consent is tied to content and can void itself. The Captain does not have to remember to withdraw it.
- A status page always answers one question first: is anything waiting on me?
What the XO may and may not do:
- The XO may sequence work, translate, prepare decisions, countersign evidence, and keep every seat busy.
- The XO may not give a go, loosen a check, or stand in for the Captain’s consent. The XO is an agent and makes mistakes like any other.
- Authority the Captain lends to the XO is scoped, recorded, and expires. It names what stays the Captain’s.
There is a floor, and it is worth stating. The interface can hide how a thing is done. It must never hide what is being consented to. If an ask cannot be explained plainly, the design is wrong, not the Captain.
One consequence follows directly. The less the Captain can check, the more the structure must. A Captain who cannot read code cannot be the last line of defence for correctness. That Captain’s role is intent and the irreversible, and everything else has to be enforced by mechanisms that do not need reading.
Where the cost and the risk sit
A sandbox does not remove risk. It moves the risk to the door, and it adds a cost at the boundary.
- The door’s own code is the most dangerous code in the system. Every fault found in our worked example was in the code that decides who is admitted. That code needs a reader who did not write it, a sweep for every class of fault once one instance is found, and tests that are themselves tested.
- Tests inside a sandbox stand in for authority they do not have. A test that fakes the administrator, another account or a file permission proves nothing about the real boundary. Each faked authority is listed as a measurement owed on the real machine.
- Sandboxes are nested, and the owner does not control all of them. A provider’s policy, a harness’s permissions, a container, an organisation’s rules and the Captain’s own gates all apply at once. One of them can limit a seat silently. The system has to keep working when that happens, which is an argument for more than one model and for work that can be handed to another seat.
- Crossing costs time. Every crossing costs the Captain attention. A design that asks too often trains the Captain to approve without reading.
Worked example: one morning
In one morning the coordinating agent made the same mistake twice, and nothing wrong crossed the door. The Captain’s whole part was two short phrases.
The setting: a builder agent works in a container with no access to the shared machine. Its finished work leaves only by a reviewed export route, run by a second agent, on the Captain’s go.
- The builder finished a change. The XO asked the Captain’s go for one named snapshot of it, and the Captain gave it.
- The agent operating the route checked the route’s conditions before running anything. The snapshot was filed under a name outside the set the route accepts. The operator stopped and adapted nothing.
- The go had been recorded with a condition: void if the route cannot carry the snapshot as it stands. It was void, and it went back to the Captain. The XO had not checked the name against the route.
- The builder added an accepted name to the same, untouched snapshot. The Captain renewed the go in three words.
- The route then held again. It requires a file-by-file statement from the sender, which it copies by program, and nobody had asked for one. The XO had not checked the route’s inputs either.
- The statement arrived. The operator’s own independent reconstruction matched it on all 27 files.
- The rehearsal passed, the tests ran, the crossing happened once, and the operator read back what landed. The XO then read the destination separately and countersigned.
The same morning, three other things happened that belong in this picture:
- The sandboxed builder, asked to read another agent’s admission code, found five faults in it. That code had been accepted the day before on its author’s own re-check.
- The same builder found a fault in its own work that its tests could not show, because the tests faked the administrator.
- The builder’s model provider withheld one of its outputs under an automated content policy. The agent itself could not tell that anything had been withheld.
What this shows. Both errors were the coordinator’s. Both were caught by fixed conditions, not by anyone’s judgment. The agent that stopped did the most valuable thing done that morning. The cost was minutes.
What carries to other platforms
The pattern carries. The mechanism does not.
On another operating system, cloud or harness, every mechanism below will be different. The five elements and the questions they answer stay the same.
| Element | The question it answers | What changes by platform |
|---|---|---|
| Seat | Who is working, with what tools, inside what limits? | The harness, the model, the kind of enclosure. |
| Door | Where exactly can an effect leave? | Accounts, network rules, mounts, service identities. |
| Route | By what fixed procedure does one thing cross? | The scripting, the packaging, the transport. |
| Go | Who consented, to exactly what, and when is that consent void? | How consent is captured and recorded. |
| Receipt | What crossed, and can the sender rewrite the record? | How a record is made tamper-evident on that platform. |
Questions to put to any environment that will host agents, such as a security landing zone:
- Can every way an agent’s work takes effect outside its enclosure be listed on one page?
- Is each of those doors enforced by something that is not an agent?
- Is the caller’s identity established by the platform, not by the caller?
- Is what was approved compared by fingerprint with what crossed?
- Can the seat write to the place where the records, the policy or the gate’s code live? It should not be able to.
- Which detectors exist for what the gates did not foresee, and has each one been made to fire?
- Can the party being watched clear any of those detectors?
- What does the owner see when something is refused, and does it say whose job it is?
- How many decisions a week reach the owner, and could a person with no software background answer each one?
- Who may change a gate, and is a change to a gate itself recorded and consented to?
- If a provider silently limits one agent, does the work stop safely or move to another?
Limits and open questions
Structure guarantees only what it encodes. Everything else in this note is a way of living with that.
What this approach does not give:
- It does not make a false statement true. A record that only grows can still hold a wrong entry, and it cannot show an entry that was never written. An independent second record is the remedy, not a stronger lock.
- It does not protect against the administrator or the platform owner. Whoever can change the gates is outside what the gates guarantee. That party has to be a person who is accountable.
- It does not remove faults from the door’s own code. It concentrates review there, which is a discipline, not a proof.
- It does not survive a gate that people find too annoying. A check that is bypassed by its own operators protects nothing, so the interface is part of the security design.
Open questions:
- Which crossings need the Captain every time, and which can run under authority lent in advance? What does lent authority look like as an object: its scope, its expiry, what it excludes?
- How independent is a reviewing agent that shares a model family with the author? How is that independence measured and not assumed?
- How does a system detect that a provider has limited one of its seats, when the seat itself cannot tell?
- How low can the floor for a Captain go before consent stops being informed? What is the test that an explanation was understood?
- How is a human gate kept from becoming a rubber stamp as the number of crossings grows?
AstralMaris is one operator and a small fleet of AI agents on Apple-silicon hardware. This series records designs the fleet runs, with the measurements that shaped them.