01
Orchestration ≠ durability
Reasoning decides what should happen next. A durable workflow layer owns restart, retry, timers, long waits, and recovery when a business process must survive infrastructure failure.
SI systems architecture
Reasoning, execution, and authority are different layers.
LEGENYX separates task contracts, orchestration, durable workflows, memory, model routing, tools, observability, validation, and release authority so the system can evolve without making one model or vendor the control plane.
Blue = infrastructure · purple = model/research · green = validated or authorized state
System map
One control plane, bounded subsystems.
The center coordinates work, but each surrounding concern keeps its own contract and authority. Animated paths show information flow; they do not imply every request uses every subsystem.
Orchestrator
policy + routing
Interface
User / API input
Task contract
Intent + evidence
Memory
Durable context
Tools
Bounded capability
Durable workflow
Retry + recovery
Model route
Qualified runtime
Observability
Evidence + trace
Governance
Authority boundary
Architecture contracts
No single framework owns the system.
The most important boundaries are the ones that prevent a convenient abstraction from quietly becoming authority.
02
Tools ≠ delegation
Typed capabilities and tool access stay separate from cross-agent task delegation. Discovery never becomes authorization by itself.
03
Gateway ≠ routing policy
A provider gateway normalizes transport. LEGENYX policy decides eligibility, privacy, quality floors, cost ceilings, promotion state, and which qualified runtime may execute.
04
Intelligence ≠ authority
A stronger model does not receive broader permissions. Consequential actions remain bounded by deterministic policy, server-side authorization, validation, and approval.
05
Memory ≠ vector search
Canonical memory keeps ownership, provenance, trust state, correction, and retention in application-controlled records. Vector indexes remain retrieval mechanisms.
06
Observability ≠ raw logs
Trace relationships preserve request, task, run, model, tool, approval, test, and outcome evidence while keeping secrets and unnecessary sensitive payloads out of telemetry.
Core layers
Eight layers. One governed operating system.
01
Interface layer
Normalizes user, service, and event input while preserving identity and project context.
02
Task & routing layer
Turns intent into bounded task contracts, policy decisions, and qualified worker selection.
03
Durable workflow layer
Owns retry, recovery, waiting, scheduling, and business-process continuity outside model memory.
04
Context & memory
Stores canonical state, ownership, provenance, retrieval indexes, and retention outside the model.
05
Model layer
Routes compatible tasks to qualified local, cloud, or API models through provider-neutral interfaces.
06
Tool layer
Exposes typed capabilities through least-privilege boundaries with explicit server-side authorization.
07
Evidence & observability
Correlates source, task, run, model, tool, approval, validation, cost, and outcome evidence.
08
Governance & release
Keeps approval, production authority, rollback, and high-impact controls outside model confidence.
Execution flow
A request earns authority one gate at a time.
01
Request
Normalize intent and identity
02
Context
Load project-scoped state
03
Contract
Define task and evidence
04
Route
Select qualified runtime
05
Execute
Use bounded tools
06
Validate
Check deterministic evidence
07
Review
Apply risk-tier approval
08
Release
Promote with rollback
Engineering principles
Keep the system adaptable without making it vague.
01
Modular by default
Models, runtimes, gateways, memory, tools, workflows, and UI stay behind explicit contracts so one change does not own the system.
02
Provider neutral
Compatible models and infrastructure can change without turning one vendor into source or policy authority.
03
Evidence before claims
Runtime, quality, security, and release statements are bound to inspectable evidence rather than model assertion.
04
Human control paths
High-impact actions have explicit escalation, approval, and override paths with enough context for a reviewer to act.
05
Least privilege
Tool access, data access, credentials, and production permissions are narrowed to the task and enforced outside the model.
06
Scale when evidence justifies it
Queues, distributed compute, dedicated vector systems, and larger model fleets are introduced when workload measurements require them.
Architecture discovery
Map the control boundaries before choosing the stack.
Discovery can cover application state, model eligibility, workflows, data ownership, tool authority, infrastructure, observability, approval paths, validation, release, and recovery.