On this page
Revision r1 · October 2026 · AstralMaris design series The thesis behind the fleet’s architecture. Every other paper in this series — the container, the door, the build receipt, the airlock, the identity pass — is one mechanism in service of the claim below.
1. The claim
Every agent-observability product answers two questions: what did my agents do? and what did it cost? Those answers converge on a commodity. The question that actually troubles the person running an autonomous fleet is a different one, and it is unanswered:
Is my fleet aligned with what I am trying to accomplish — and where it is not, or where I cannot tell, do I know that I cannot tell?
What troubles people about autonomous agents is not error. It is not-knowing: the serene dashboard over a dark fleet, the green check that means “process alive” rather than “doing the right thing”. So the differentiator is not more metrics. It is a layer that is honest about its own uncertainty — one that says what it knows, what it infers, and what it cannot see, in a form a time-limited human can hold in their head.
That is the inverse of the industry reflex, which is to manufacture confidence.
2. Why it is hard: the autonomy paradox
Monitoring is hardest exactly when the agents control the direction of the work. Stated precisely: the moment the overseen party authors the standard it is judged against, every downstream check becomes circular. Any alignment metric the agents help produce measures the agents against their own frame.
Three independent literatures — agency economics, AI control, institutional governance — converge on one answer: anchor the yardstick on something the agents cannot author. Four such anchors recur:
- Frozen, human-authored intent the agents cannot rewrite — a commander’s intent; the principal owning the objective.
- A trusted-but-weaker checker in the verification path, so the evaluator is never the evaluated.
- Behavioural tripwires that fire on observed action, never on self-report. Whatever behaviour is observed, a capable agent may be acting aligned in order to pass; so nothing is anchored on what the agent says about itself.
- Adversarial cross-agent review, so a weaker judge can adjudicate work it could not have produced.
And one failure mode to instrument against above all: the rubber stamp — the overseer who formally approves but no longer meaningfully reviews. Its measurable tell is an approval rate trending to one hundred percent while review time shrinks.
The design consequence: semantic governance cannot be a score the agents help compute. It is a legible account of the fleet’s activity, measured against frozen human intent, plus tripwires and drift proxies the agents do not author — surfaced with its blind spots shown.
3. What to observe, at what altitude
- Process, not outcome. Outcome-only scoring is blind to an agent that reaches the right answer for the wrong reasons, which is the signature of incipient misalignment. The reasoning trajectory is scored against intent, not the deliverable’s acceptance. A portable finding from the drift literature: drift happens mostly through inaction — the quiet decay a metrics wall never shows.
- The causal path, not every step. Per-step monitoring drowns the human; alarm floods are a known failure of control rooms. Show the chain that actually gated the result; mine the event log into a process map and narrate deviation from the intended plan; manage by exception with an explicit attention budget.
- A friction ledger, not a score. Correction events, intent staleness, reversal and rework rate, dissent rate, time to consensus, overseer approval rate with review time, agent-authored specification changes. Regression is the trend in that friction. A model asked “is this aligned?” is one input signal, never the metric — judging a model’s alignment with a model is the circularity trap.
4. The legibility layer
The job is to turn a fast, multi-agent firehose into a picture a single time-limited human can comprehend and trust, because it is honest about what it does not know.
- Three levels, as in situational-awareness doctrine: perception (data), comprehension (“the fleet is converging on a wrong decision”), projection (“if this continues, the merge fails in three steps”). A trace tree stops at the first. The product delivers the second and third, each tagged with confidence.
- Finished intelligence, not raw collection. The human receives an assessment, a projection and a confidence, with the raw trace as drill-down, never as the default view.
- Pre-declared must-knows. Before the fleet runs, the human declares a small set of trip conditions — an agent stalled, looping, out of scope; the work drifting off intent. This pre-commits attention and makes omission detectable: a declared must-know cannot be quietly dropped from the narrative, because it is checked for explicitly.
- Narrate as intent versus actual. Activity is legible only against a declared end state. The story is “the intent was X; here is how each agent advances or diverges from X”, not “here is what occurred”.
- Honest uncertainty. Negative evidence is structurally unavailable to passive observation — “idle” and “dead” look the same without an out-of-band signal — so the layer renders confidence bands, never binary health. Generated narratives are known to omit critical facts a large fraction of the time, and omission is worse than fabrication for oversight; so every claim in a narrative is verified against the trace, grounded in the declared must-knows, and drill-down-able. The narrative is the top layer; the trace is the auditable substrate; the model is never the sole source of truth.
This narrative layer over standardised agent traces is the open whitespace.
5. The fleet as the map
The fleet that wrote these papers already runs a primitive of every mechanism above, which is the sense in which its own operational history is the reference design:
| mechanism | what the fleet runs |
|---|---|
| frozen exogenous anchor | standing orders the agents cannot rewrite: only the operator merges to main; no live change without a go; a go is quoted with the question it answers |
| correction-event ledger | a lessons register, each entry with the evidence that produced it |
| intent currency | handoffs written at every gate crossed, read at every cold start; a stale handoff is a flag |
| adversarial cross-agent verification | independent readers from a different model family; two native gates on one piece when it is worth the cost |
| tripwires not authored by the actor | the entrypoint assertions, the collector’s refusal, the leak scan, the typed-outcome monitor rows |
| rubber-stamp avoidance | the operator’s rulings are recorded in his own words, with what he reversed |
The product is, literally: instrument the governance loop the fleet already lives, anchor it on intent the agents cannot author, and surface it as honest, finished-intelligence deviation from intent.
6. New words
Borrowed words failed the fleet: “dry run”, “ready”, “pin” each meant different things to different seats until measured. The operator’s answer is not an ontology — meaning here is situational and adjusts to prevailing conditions — but a registry of practice tied to compression points: the places where a crew’s shared understanding is compressed into a word, and where a word that has drifted from its practice is the first sign that the crew has. Sensing that drift is the same job as sensing alignment drift, one level down; indexing the registry by stakes and by how far a word has detached from what it names is the open problem this series will return to.
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.