Live orchestration · The coordination surface

Guardian One OS

A fleet of autonomous agents, one orchestrator, one egress gateway — the coordination surface where a system runs itself and shows its work.

What it is

The control plane for an autonomous agent fleet.

Every agent declares its scope. Every external call goes through one hardened gateway. Every decision is logged with evidence and snapshotted before it runs. This is the surface where orchestration becomes observable — and reversible.

The architecture

Coordination you can inspect, audit and undo.

01

One orchestrator

A single control plane resolves cross-agent conflicts and keeps a human in the loop. No agent acts alone; none drift.

02

Hardened egress

Every external call leaves through one gateway — TLS enforced, rate-limited, circuit-broken. One audited door, not many.

03

Evidence & snapshots

Every decision logs structured evidence; every mutation snapshots prior state before it runs. Explainable and reversible by default.

04

Local-first reasoning

Inference runs on-prem first; the cloud is an explicit, logged fallback — never the silent default.

The decision path

How a decision flows.

From intent to action, every step passes through the same auditable path — and every step can be walked back.

The control plane — fleet, orchestrator, gateway, evidence, snapshot, revertAgent fleetOrchestratorEgress gatewayWorldEvidence logSnapshot storerevert
01

Intent

A goal enters the system — from the owner or a scheduled trigger.

02

Orchestration

The coordinator picks the agent, checks its scope, and sequences the work.

03

Egress

Any outside call exits through the one hardened gateway — TLS, rate-limited, circuit-broken.

04

Evidence

Inputs, reasoning, alternatives and outcome are logged before the result lands.

05

Snapshot

Prior state is captured so the change is reversible by construction.

06

Action — or revert

The mutation runs; if it's wrong, revert restores the snapshot deterministically.

The fleet

Dozens of agents. Four runtimes. One registry.

Every agent in the system must exist as a row in a single version-controlled registry, enforced in continuous integration — code that isn't catalogued doesn't ship. Specialists, not a monolith:

01

Operations & observability

Dedicated triage, drift, monitoring and cost agents watch the system itself — regressions, anomalies, spend.

02

Finance

A cfo agent, bookkeeping and a spending circuit-breaker that trips on anomalous outflow before it compounds.

03

Security

A varys sentinel on watch — detection rules, anomaly scoring, credential hygiene, containment playbooks.

04

Communications

A hermes messenger and email-triage agents route what matters and quarantine what doesn't. Outbound stays gated.

05

Knowledge & memory

An archivist and a verifier keep records, provenance and claims honest — every fact carries its evidence.

06

Domain platforms

Deep vertical agents for professional workflows — licensing, credentialing, relocation — built on the same spine.

Autonomy, governed

Free to act — inside a ladder of stakes.

Every action is classified by what it would cost to be wrong. The system proposes; the owner disposes. No agent merges, spends or emails on its own.

01

Auto

Reversible and cheap — code, tests, docs. The agent acts silently and the ledger records it.

02

Notify

Reversible, inside the system — drafts and branches. The agent acts, and it rolls up in the digest.

03

Stage

Costs money or leaves the system. Queued for explicit owner approval before anything moves.

04

Halt

Irreversible or out of bounds. Hard stop, immediate alert, no override from inside the fleet.

The loop

Every mistake becomes a rule.

The fleet keeps a working memory of its own errors and wins — not as logs to grep, but as rules it consults before acting again.

01

Evidence

Each decision logs its inputs, reasoning and alternatives — before the outcome lands, not after.

02

Recall

Agents retrieve relevant prior lessons at decision time, so the same mistake doesn't get a second run.

03

Supersede

Stale or wrong lessons are retired explicitly. Memory stays curated, not accumulated.

Threat model & rollback

Reversible by construction.

Agents are scoped, egress is single-doored, and every change is snapshotted before it runs — so a bad decision is contained and undoable, not a cleanup.

01

Scoped agents

Each agent declares the resources it may touch; nothing is ambient.

02

Snapshot + revert

State is captured before mutation; revert restores it without guesswork.

03

Notify on failure

Errors escalate to a human through the active visibility channel — never swallowed.

04

One audited door out

All outbound traffic exits through a single gateway — TLS-enforced, rate-limited, circuit-broken.

05

Kill switch

A fleet-wide halt control stops every autonomous loop at once. Built in from day one — and used.

The fleet, live

Divergence-free flow. No agent drifts.

The field is curl-noise — mathematically divergence-free, so nothing leaks or pools. Drive it with your cursor: the same property that keeps the fleet coordinated under load.

The mandate

Protect. Enhance. Empower. Grow. Every action explained. Every mutation reversible. No agent unaccounted for.

One orchestrator. Full accountability.

Guardian One is in active development as a coordination surface for autonomous agents. Request the system brief for architecture, threat model and the audit/rollback design.

Request the system brief →