Agentic Systems · AVSFT

Agentic runtime with authority attached.

AVSFT is the operating protocol for governed agents — persistent daemons with durable identity, explicit authority, tool permissions, reconstructible state, receipted execution and human-bounded decision support.

The daemon is more than the model.

AVSFT separates daemon identity, runtime state, authority, tools, memory and model cognition so changing one component does not silently redefine the actor. The architecture continues a systems lineage that predates large language models: feedback and control, concurrent state, long-running processes, supervised actors, durable event records and recoverable services.

Pre-LLM systems lineage

The design begins with the late-1940s control and communication tradition and continues through Petri and workflow models of concurrent state and transition.It incorporates long-running daemons, actor isolation, Erlang/OTP-style supervision, statecharts, virtual actors, event sourcing and durable execution. AVSFT applies that lineage to persistent agent operation instead of treating a chat session as the system boundary.

Resident daemons

A resident system is a long-running daemon with its own lifecycle, identity, local state, mailbox, timers, capabilities, authority and recovery contract.Model cognition can be attached, replaced or upgraded without renaming the daemon, erasing its history or silently expanding what it may do. Individual resident identities remain private implementation detail.

Mechanisms retained

AVSFT retains typed asynchronous messages, bounded restart topology and escalation, append-only transition history with snapshots, process-independent logical identity, event-triggered cognition, evidence ancestry, active-belief freshness and explicit nondeterminism boundaries.Petri and workflow nets are used as verification projections; they do not replace the production runtime.

Mechanisms bounded or rejected

Shared working state may coordinate processes but cannot become an untyped global truth store.Delegation may consider capability and availability but cannot auction away authority. Repeated summaries do not become independent evidence, and model-owned memory does not become institutional identity. A fixed model judge cannot be the sole verifier, and replay cannot silently reissue external side effects.

Identity & continuity

Persistent actor identity remains distinct from whichever model, provider or session is currently attached.That lets cognition change without silently replacing the operating actor or its authority history. Behavioral continuity across model swaps remains an empirical claim until the paired runtime tests pass.

Event-triggered cognition

Deterministic sensing and state continuity remain resident; expensive cognition wakes when its expected decision improvement justifies compute, latency and opportunity cost.Novelty, uncertainty, disagreement, consequence, staleness, ambiguity, authority boundaries and verification deficits are candidate wake signals that still require workload calibration.

Evidence ancestry and memory

Confidence must follow independent evidence roots rather than the number of messages, summaries or agents repeating them.Immutable custody history remains intact while active beliefs may expire, be superseded or require revalidation according to their class and observed change rate.

Supervision & recovery

Heartbeat, health, acknowledgment, bounded restart intensity, escalation, cancellation, replay and recovery are runtime responsibilities.Interrupted consequential work becomes UNKNOWN until reconciled; a restarted process cannot invent success, failure or continuity from topology alone.

Authority & commitments

Tools and actions execute inside declared jurisdiction and approval boundaries.An agent can prepare negotiation options, conditions, counterpositions and contract-review packets, but a commercial or legal commitment remains subject to the human or owner-system authority assigned to that decision.

Adversarial decision support

High-consequence choices can be structured around evidence, realistic alternatives, bargaining position, enforceability, reversibility, precedent, timing and downside.The system can preserve competing views and identify what evidence would reverse a recommendation instead of allowing one generated answer to grade its own work.

Receipts & reconstruction

Material transitions, evidence references, approvals and release states can emit receipts so the path from observation to recommendation to authorized action remains inspectable.Exact replay consumes admitted records; nondeterministic model or tool work requires an explicit captured activity boundary.

Five functional roles

AVSFT is a five-role agent substrate organized around operations, governance and policy, communications, research, and assurance with anomaly detection.The roles collaborate through typed state and receipts while retaining separate authority, evidence and recovery responsibilities. The public model names the functions rather than the private resident identities.

RTC and RTG trace context

AVSFT can consume Reach-owned trace context through a bounded connector.RTC records the context of one arriving object; RTG preserves the temporal graph of recorded journeys, handshakes and traversals. Trace recurrence can guide investigation and orchestration, but it does not prove identity, intent, truth or causation.

Operating Model

Five roles. One governed substrate.

Each lane owns a distinct responsibility. Trace context crosses the system through bounded, receipted interfaces.

RTCContext at arrivalReach-owned boundary scan
RTGThe road recordedReach-owned trace graph
AVSFTGoverned responseRole-specific action and receipt

HOW WE AUDIT SYSTEMS

AuthorityBind the source
StateTrace the system
FailureTest recovery
CutoverProve it live

The sequence distinguishes what is documented, implemented, deployed and proven. A later state never substitutes for missing evidence in an earlier one.

Continuity

Persistent operation should remain inspectable.

AVSFT preserves identity, authority, state and evidence across long-running work. Teams can change models, recover interrupted processes and review consequential actions without losing the operating history that explains what happened.