The twelve canonical layers

Human Interface is the layer where a person meets the system: it captures human intent, carries approval requests to a reviewer and back, and gives operators a surface to watch what a governed multi-agent system is doing. Dashboards, admin panels, operator APIs, and any command UI a human starts, reviews, or blocks work through, not the workflow logic that acts on a decision, and not the policy logic that decides whether an approval is required in the first place.

Goal & Intent is the layer where a business objective becomes something the rest of the system can act on: a structured goal schema, an objective definition, a playbook, or a KPI target, not a paragraph of prose or a verbal instruction. Its job is translation, not execution.

Orchestration routes work, decomposes goals, sequences tasks, and manages retries and approvals. It decides what happens next, in what order, on what trigger, and with what retry and escalation behaviour, without performing domain reasoning or holding long-lived memory of its own.

Agent Runtime executes agents with bounded scope and lifecycle management. It owns agent lifecycle, invocation context, execution state, and communication patterns: the mechanics of how a single agent-driven step actually runs and returns a result, distinct from Orchestration's concern of what happens next across a whole workflow.

Agent Registry defines discoverable capabilities and boundaries. It is the shared discovery, versioning, and lifecycle surface for every agent on Synedrix OS: answering what agents exist, what they're permitted to do, and whether a given agent is currently eligible to be invoked, as distinct from Agent Runtime, which actually executes an invocation once the registry has cleared it.

Tool Abstraction normalises how an agent reaches an external system. Every capability an agent may invoke (a database read, a write to canonical state, a search call, a cross-agent delegation) passes through this layer as a named, typed tool rather than a direct integration the agent's reasoning code holds a handle to.

Memory & Data persists business state and contextual knowledge, and owns the memory control plane: the governed process by which a record is formed, consolidated, promoted, recalled, decayed, and erased. Five self-hosted stores answer where a byte lives: Redis for hot state, Supabase PG17 for canonical relational truth, QuestDB for time-series, Neo4j for the context graph, Qdrant for semantic vectors. A lifecycle governs how a record moves between them.

Event standardises how a fact enters, travels through, and leaves Synedrix OS. Signals, approvals, state changes, failures, anomalies, and callbacks all propagate as events, and every one of them uses the same envelope regardless of what emitted it.

Governance & Controls enforces policy, permissions, approvals, and compliance. It decides whether a proposed action is allowed to happen, not whether an agent is capable of doing it, and not what the action itself accomplishes, which this layer must never absorb.

Observability makes execution measurable, explainable, and reviewable (for every workflow, every agent invocation, every tool call, and every memory operation) without deciding whether an action was permitted or owning the business state it touched.

The Domain Pack layer is where Synedrix OS stops being domain-agnostic. Everything below it (orchestration, agent runtime, registry, tools, memory, events, governance, observability) is reusable across any business domain. A domain pack injects the domain-specific entities, workflows, agents, tools, policies, and KPIs that make the platform do something for one business: Trading, Commerce, Ops, Finance.

Business Instance configures a concrete runtime. Everything above it (the OS layers and whichever domain packs are selected) is a reusable definition; this layer is where that definition becomes one customer's actual deployment: which packs are enabled, what credentials it holds, what thresholds and approvers apply, what deployment mode it runs under.