▸ Browse docs — Core concepts
▸ SECTION B · Core concepts
Nodes
Last updated: 2026-08-16
A node is one step in a graph. The 19 node types:
Execution
| Node | Purpose |
|---|---|
agent | A model-backed agent step. Config: model, system prompt, temperature, tools. |
tool | An MCP tool invocation. Every call is a policy decision point. |
transform | Pure data transformation between steps. |
router | Conditional branch on a field or classifier output. |
terminate | Ends the run. |
trigger | Starts the run: inbound email, webhook, schedule, event. |
Gates and human work
| Node | Purpose |
|---|---|
gate.hitl | Human approval with quorum, SLA, escalation. |
gate.policy | Policy evaluation gate (in addition to edge-level policy). |
gate.eval | Evaluation-suite gate — blocks until the suite passes. |
human_task | An assignment to a human (not an approval — work). |
Structure
| Node | Purpose |
|---|---|
map | Fan-out over a list; each item runs the subgraph. |
subgraph | Nested graph reference. |
wait | Pause for a duration or a condition. |
Memory
| Node | Purpose |
|---|---|
memory.read | Read from the workspace memory store. |
memory.write | Write to the workspace memory store (a ledger event). |
rag.retrieve | Retrieval from a connected vector store. |
Example config
nodes:
triage:
type: agent
model: anthropic/claude-sonnet
tools: [search, classify]
approve:
type: gate.hitl
quorum: 2
roles: [ops_lead, compliance]
sla: 4h
escalate: vp_ops
Every node type is covered in the DSL reference with its full schema.
Was this helpful?