Architecture Overview
Three layers, one record: how a governed agent fleet is put together
AstralMaris is a fleet of AI agents run by one human. Everything it does passes through
three layers that are deliberately kept apart: a hub that keeps the record, supervisors that
decide when an agent works and hold it to a budget, and the seats where the agents actually
think. A fourth party, independent verification, reads all three and keeps an overall
message state awareness and provenance through deterministic mechanisms. This page is the
map; the tables underneath it define every interface and every gate in plain words.
The major principle the whole design follows:
- every rule is a deterministic structural check the hub makes when a message is sent.
- a supervisor confirms when an agent wakes and the rule conditions are verified by an independent instrument.
- an agent's judgement is an input that arrives as a cited receipt, never the end state of a rule.
Click any numbered marker in the drawing to jump to its definition.
On smaller screens, scroll the drawing horizontally. Its numbered definitions follow below.
The three layers, and why they stay apart
The hub is the record. Ming-Qiao stores threads, messages, decisions and artifacts,
and applies the rules that can be checked at the moment a message is sent: who is allowed to speak
as whom, whether the envelope is well formed, whether a sender is flooding, whether the sender holds
the lease for the work it claims. It knows what arrived. It does not decide what to do about it,
and it never starts a process. That last point is a security decision, not a stylistic one: a component
that both admits messages and launches agents is a component whose compromise moves the whole fleet.
Keeping the hub to the record means a compromised hub yields a corrupted record, which the verification
column exists to catch, and nothing moves until a supervisor with its own provenance chain decides
it should.
The supervisors are the mechanism. The Awakener is the name of the programme that
replaces a human typing "check" with a bounded, machine-triggered wake. A supervisor subscribes
to the hub, decides that an event warrants a wake and in what shape (a quick check, a long piece
of work, a review, or a "ruling needed" that goes to a human), hands the seat exactly one current-state
record and one budget, holds it to that budget, records a receipt for every decision before acknowledging
anything, and retires the seat completely when the work is done. Supervisors are the unit that acts,
so they are the unit that carries the pause switch and the kill switch. There is one per seat or
per host, never one for the fleet.
The seats are the cognition. A seat is built around a harness — the agent program
that calls a model and acts on its answers with its own tools — and a governed identity. The model
is one declared field of the harness profile. It runs only when woken, with a credential that exists
in the hub's registry as a digest and nowhere in plaintext it could read, and with a guard on what
it may emit that runs before anything is sent. Its judgement is valuable and is treated as an input:
it arrives at the hub as a cited receipt or an artifact by hash, and a rule is never satisfied by
a seat saying it was.
Verification reads everything and writes nothing. An inventory of record says what
should be running, by digest. A provenance gate compares that to what is running and prints both
values, so the reader knows which side moved. An independent checker runs outside the identity the
agents share and outside the author's hand. A review lane reproduces every claimed number and shows
every control failing before it is believed. None of this is optional in a fleet that acts without
a human trigger; it is what makes the trigger safe to remove.
Interface points
Security gates
A gate is a check that can refuse. Each one is engaged in reporting mode on a
development host (it prints what it would have refused) and in enforcing mode once the
system is promoted to a container. A gate that has never been seen to fail is not a gate; every
one below has a known input that makes it refuse, and that refusal is demonstrated before the gate's
pass is trusted.
Glossary
- Hub
- The message and record service (Ming-Qiao). Threads, messages, decisions, artifacts, lease authority. It never launches anything.
- Supervisor
- The process that decides when a seat wakes, hands it one record and one budget, records receipts, and retires it. The Awakener programme is the design and delivery of supervisors and their gates.
- Seat
- A confined working position for a harness that calls a model and acts through its tools, with a governed identity and a credential known to the hub only by digest.
- Wake
- One bounded episode of agent work, started by a supervisor on a hub event, never by the agent itself and never (once the Awakener is live) by a human typing "check".
- Current-state record
- A versioned artifact, produced by the hub at a boundary such as the day's end, that says where each thread stands. A waking seat is handed one of these instead of a thread to reread.
- Receipt
- A durable, deterministic trace of one decision, written before that decision is acknowledged. Receipts are how the record can be reconstructed from the store alone.
- Lease and epoch
- A lease says who holds a thread's work right now; the epoch is the number that proves which grant of the lease a message belongs to.
- Inventory of record
- The typed list of everything that should be running, each with its expected digest, kept in version control and mirrored as a root-owned copy the agents cannot edit.
- Provenance gate
- The instrument that compares what is running to the inventory, prints both values, and says which side moved.
- Red before green
- The discipline that every control is first driven against a known failure and seen to refuse, and only then trusted when it passes.
- Reporting and enforcing
- The two modes of every gate. On a development host a gate reports what it would refuse; in a container it refuses. Nothing is ever silently off.