Execution model
The executor is stateless per request — every invocation reconstructs state from PostgreSQL. A long conversation runs as many short executor calls, driven by inbound events (transcripts, tool callbacks, child-flow completions). This makes scale-out straightforward: any backend instance can pick up the next event.Node types
15 node types organised by purpose:Core (9)
Core (9)
LLM (1)
LLM (1)
Tools & knowledge (3)
Tools & knowledge (3)
Composition (2)
Composition (2)
Variables
Theagentic node extracts typed variables from the conversation:
string, number, boolean, date, phone, email, and image.
- Asked vs derived — a variable with a prompt is asked of the user; a variable marked derived is authored by the assistant from context and never asked (e.g. a ticket subject line).
- Image variables are filled by the runtime from media the user sends — never by the LLM.
- Persistent variables survive across sessions for the same user.
- Whether a variable is required is decided per edge: an edge can require variables (or a photo on the current turn) before it can be traversed.
Parallel execution
Parallelism is expressed on the graph itself: multiple edges leaving the same port, marked parallel, fan out into concurrent branches. The engine computes where the branches reconverge and merges their results. Branches coordinate at runtime: only one branch speaks to the user at a time (the first to produce output wins the voice), andwait nodes
in sibling branches cancel automatically when another branch starts
responding.
Invoke → return (child flows)
A parent flow can invoke a child flow — the pattern behind consultations, relays, and warm transfers:- Ringing — a dial target is a person, not a device: ringing reaches all their endpoints (in-app notification on every signed-in device + their phone number); the first answer wins. Multiple targets can ring in parallel or sequentially.
- Hold + narration — the original caller can be held with a message or music, and the parent can narrate child progress (“I’m reaching your advisor now…”) driven by child lifecycle events.
- Completion types —
return(the agent comes back to the caller with a result),bridge(the caller is connected directly to the consulted party), orhandoff(external transfer). - Keep-alive — a child can be kept alive after returning, so a follow-up invoke reconnects the same consulted party without ringing again.
Tools
ThetoolExecutor node runs configured tools:
Tools are authored directly, imported from a cURL command, or
imported from an OpenAPI spec. Tool secrets are scoped per
organization and injected at invocation time — flow definitions never
contain credentials.
Asynchronous tools: a tool that reports async completion either
routes the flow onward immediately on a pending port or parks the
session; your system resumes it later by calling a single-use,
HMAC-signed callback URL.
Memory — the brain
One knowledge graph per organization. Every person has a personal brain, the organization has a company brain; both live in the same graph, separated by scope.- Episodes — the transcript is kept verbatim as evidence; every fact links back to the conversation it came from.
- Statements — atomic facts about people, assets, locations, processes: subject, predicate, object or value, with the time they held and who asserted them. A new value for a one-per-subject predicate supersedes the old one; nothing is silently overwritten.
- Scope — facts tagged personal stay with the person; facts the
flow allows to be learned (
autoLearn) enter the company brain. A fact inferred about a colleague is shown to that colleague first. - Feeds — external sources (a scanner, a MES, a phone’s location) push events that map to live state without a model in the loop.
- Persistent variables — flow variables marked persistent are restored at the start of the user’s next session.
- Knowledge bases — per-organization document collections;
ingestion chunks and embeds, retrieval through the
ragnode.