Skip to main content
A flow is a graph of nodes. Knowing what the set contains is the fastest way to know what a flow can and cannot do — because if a behaviour is not in this list, a flow cannot do it, and it belongs to the AI agent instead.
Seven entries are buildable on the canvas today, plus a composite shortcut. The engine executes the ten node types documented on this page. Of those, send_message, send_amb_primitive, wait_for_reply, branch_on_reply, and assign_to_agent appear in the builder palette, alongside two entries that aren’t part of that documented nine: end, and integration — six entries, one per connected app, all coming soon (see Canvas basics for what they do today, which is nothing). The canvas also offers a “Message with options” shortcut that isn’t a node type of its own — it drops in a preconfigured send_message paired with a branch_on_reply whose fallback is hardcoded to loop with max_retries: 2.set_conversation_attribute, add_label, conditional, delay and answer_opening_question are fully implemented in the engine and run correctly — they simply cannot be added from the canvas yet. They are documented below because a flow created through the API can use them, and because they are what the canvas will gain first.

​
Sending

​
plain text out
Sends text. Supports Liquid templating, so you can interpolate captured variables and contact attributes rather than sending a fixed string.
​
rich Apple message out
Sends an Apple Messages for Business primitive. Configured with a primitive_type and the primitive’s own payload. See Channel-native nodes.

​
Waiting and branching

​
pause
Pauses until the customer responds. Flow execution state is persisted, so a flow can wait as long as it needs to — this is not an open connection held for hours.The timeout_seconds field is coming soon. The builder lets you set it, but nothing reads it today — there’s no timeout job. A wait persists indefinitely until the customer replies, however long that takes, regardless of what you enter here.
​
route on what they said
Routes on the customer’s reply using match clauses, with a fallback for anything that matches nothing. The fallback can loop back or escalate.The fallback is the important part. Customers say things you did not anticipate; a branch without a sensible fallback is where flows strand people.
​
if/else on state
Branches on variables rather than on a reply — what you already know, not what they just said. Configured with a condition and true/false next nodes.
​
timed pause
Waits a fixed duration_seconds before continuing. Use sparingly — a deliberate pause can make a bot feel less robotic, but it is also just latency the customer experiences.

​
Changing state

The two nodes in this section — set_conversation_attribute and add_label — run correctly in the engine but are not yet in the builder palette. A flow authored through the API can use them.
​
write a value
Sets a key to a value on the conversation, with Liquid support in the value. This writes a conversation attribute — not an entry in the execution’s variable bag.Reading it back through {{ variables.<key> }} is coming soon. Liquid’s variables scope is populated only by last_reply, the trigger, and two internal counters the engine writes — nothing set_conversation_attribute writes reaches it yet, so a later node’s {{ variables.<key> }} will not resolve a value you captured here. See Captured variables and Liquid for exactly what variables exposes today.
​
tag it
Applies a label to the conversation. Labels feed automation and reporting, so this is how a flow hands context to everything downstream.

​
Answering

​
hand the opening message to the AI
Hands the customer’s opening message to the AI agent once, so a flow can welcome a customer and still answer the question they started with. A bare greeting such as “hi” gets nothing extra, and an answer leaves the flow running.It only acts on Apple Messages conversations where the account setting answer_opening_question is on and the AI agent is available; otherwise it passes through and the flow carries on. It is backend-only for now, with no builder UI, so a flow created through the API is the only way to use it.Place it right before the first wait_for_reply. The answer is sent after the messages the flow already sent, but the nodes after this one are not held back while the AI works. If the AI escalates to a person, the flow ends. No flow starts while a person owns the conversation, so after the handoff the customer’s next message does not start this flow or any other. If the conversation is resolved and the customer later writes again, the conversation is no longer person-owned, so a flow can start again (an initial flow only if it is marked to restart on reopen). A takeover that is still active keeps it person-owned. The customer’s opening message is not treated as a reply to the flow’s own question, so it does not feed branch_on_reply.

​
Ending

​
hand to a human
Escalates. Three strategies: round_robin_inbox, specific_agent, or team.round_robin_inbox is coming soon. It’s a no-op stub today — selecting it sets the assignee to nobody. Only specific_agent and team actually assign an agent — use one of those two until round robin ships.Once a human messages the customer, the resolution is not billable at all — see AI-to-human handoff.
A flow also terminates implicitly, either completed or escalated.

​
Not yet available

There are no outbound-webhook or HTTP-fetch nodes. A flow cannot currently call an external service, either to push data out or to fetch data in mid-flow. Both were specified and neither is built.If your flow needs to read from or write to an external system, that work belongs to the AI agent, which reaches connected systems through integrations — and which is what determines the resolution class in any case.
This is the single most common thing people expect a flow to do that it cannot. Worth knowing before you design around it.

​
Choosing between a flow and the agent

​
Next

Branching and variables

Conditions, captured variables and Liquid.

Channel-native nodes

Which AMB primitives a flow can emit.

Testing and simulation

Running the graph before a customer does.

AI agents

Everything the node set deliberately does not cover.