Tip: select any text on this page to bring up the Ask tooltip and question the Synedrix OS assistant about it.
What Agent Runtime is
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.
This layer didn't exist as a dedicated platform until ADR-OS-001: n8n filled it implicitly, with cyclic or multi-agent reasoning simulated through hand-managed loop state. LangGraph was assigned here because its native model (stateful graphs, checkpointing, human-in-the-loop interrupts) maps directly onto the OS's stateful and cyclic execution patterns.
What it owns, and what it does not
Every layer in the architecture has a non-overlapping responsibility. The boundary is as much a part of the design as the capability.
Owns
- Agent lifecycle: how an agent-driven step is instantiated, executed, and terminated
- Invocation context: inputs, resolved tool set, and deployment mode
- Execution state: graph state, checkpoints, working memory during a run
- Communication patterns: fan-out, supervisor delegation, generator/critic loops
Does not own
- UI concerns
- → Human Interface Layer
- Business-specific workflow routing
- → Orchestration Layer (n8n)
- Policy decisions
- → Policy Guard Agent (os-agt-002)
- Tool authorization resolution
- → Orchestration Layer, which resolves and passes in the active tool set
- LLM tracing, prompts, evaluations
- → Langfuse (Layer 10)
- Canonical business state
- → Supabase PG17 (Layer 7)
Interfaces with
The other layers this one exchanges work, state, or decisions with, and why.
Orchestration
n8n invokes this layer as a governed step, carrying correlation_id, causation_id, and the resolved active tool set, and receiving back the canonical Outcome Envelope.
Memory
graph state and a checkpointer provide in-session working memory; durable recall is always an explicit, scoped tool call.
Tools
LangChain-bound tools are the runtime's only path to external systems; the executor rejects any call outside the passed-in active tool set.
Observability
every invocation emits Langfuse traces and returns an Outcome Envelope, exactly as observable as a run entirely inside n8n.
How Synedrix OS implements it
The LangGraph Runtime Implementation Guide fixes the mapping from OS concepts to LangGraph constructs: an agent-driven step becomes a compiled StateGraph, and the four cyclic/multi-agent patterns become specific topologies: iterative loops back to the agent node, reflect-and-revise pairs a generator with a critic, consensus fans out to voter nodes reduced by a quorum function, hierarchical delegates to sub-graphs. HITL approval is modelled as a LangGraph interrupt at a checkpoint.
The invocation contract is fixed in both directions: the graph never mints a new correlation_id, the executor is initialized only with the passed-in active tool set, and deployment_mode: air-gapped forces local model bindings only. Moving an agent between n8n and LangGraph is a deployment decision, not a re-registration. Its agent_id and name stay stable across either engine.
Platform mapping
LangGraph
Stateful/cyclic/multi-agent reasoning runtime, built with the LangChain library, invoked by n8n as a governed step.
OS agents at this layer
General Chat Agent
intelligence · os-agt-009Answers questions about Synedrix OS from a semantic knowledge base, with cited sources. It powers the assistant on this site.
Layer principles that apply
- Controls before execution: the governance bridge (Policy Guard reachability, HITL interrupts, mandatory Outcome Envelope returns) keeps this layer from becoming an ungoverned execution path.
- Memory is shared, access is scoped: working memory lives in graph state; canonical and semantic recall are explicit, scoped tool calls.
Agent Runtime questions, answered
Short, citable answers drawn from this page.
What is the Agent Runtime layer in Synedrix OS?
Agent Runtime is layer 4 of the twelve-layer Synedrix OS architecture, in the "Orchestration core" group. Where reasoning actually happens. It loops, accumulates state, and resumes.
What does the Agent Runtime layer own?
The Agent Runtime layer owns: Agent lifecycle: how an agent-driven step is instantiated, executed, and terminated; Invocation context: inputs, resolved tool set, and deployment mode; Execution state: graph state, checkpoints, working memory during a run; Communication patterns: fan-out, supervisor delegation, generator/critic loops.
What does the Agent Runtime layer not own?
The Agent Runtime layer does not own: UI concerns (owned by Human Interface Layer); Business-specific workflow routing (owned by Orchestration Layer (n8n)); Policy decisions (owned by Policy Guard Agent (os-agt-002)); Tool authorization resolution (owned by Orchestration Layer, which resolves and passes in the active tool set); LLM tracing, prompts, evaluations (owned by Langfuse (Layer 10)); Canonical business state (owned by Supabase PG17 (Layer 7)).
The full reference page, with every canonical specification this layer touches, lives in the documentation repository, which is private while the platform is being built.