​

Publishing & Going Live

Building a workflow on the canvas only creates a draft. To make your agent actually follow it on real phone calls, you validate it, publish it, and (when you want) test it from the browser. This page walks through each toolbar action and then recaps exactly what happens on the wire once a workflow is live.

​

The path to live

​

Validate

Validate runs the engine validator and surfaces the results in the Issues tab. Issues come in two severities: Common things the validator flags:

• Dead-ends — a node that the flow can reach but can never leave (no outgoing edge and not an End Call). The call would get stuck on that phase.
• Undeclared variables — a {{variable}} referenced in a prompt, edge equation, or Set Variable template that isn’t declared in the Variables panel.
• An unconfigured API Request — an API Request node with no endpoint, webhook, or free-form request configured.

​

Save vs. Publish

These are two different actions, and the difference matters: Only the published version runs live. You can keep editing and saving drafts while an older published version continues to answer calls; your in-progress edits don’t reach callers until you Publish again.

​

Enabled / Disabled toggle

The toolbar’s Enabled / Disabled toggle is the fast off-switch:

• Enabled — calls follow the published workflow.
• Disabled — calls ignore the workflow and fall back to the normal agent (its Behavior-panel prompt and tools), without unpublishing.

Use Disabled when you want to quickly turn the workflow off — during an incident, while you rework a phase, or to A/B against the plain agent — and flip it back to Enabled when you’re ready, with no re-publish needed.

​

Test call

Test call launches a web call from the studio so you can talk to the workflow from your browser — no phone needed. Connect your mic, then walk the flow: trigger each branch, confirm collectors capture cleanly, and watch globals interrupt and return.

​

Choosing the Director model

The Director is the fast LLM that runs on ambiguous caller turns to extract the current node’s variables and pick which prompt-condition edge is met. You set its model and temperature in the Settings tab. The Director is gated: it only runs when routing is genuinely ambiguous. Deterministic equation edges and global enter-conditions are evaluated in code without it, which keeps latency low on most turns.

​

How it runs on a call

Once a workflow is published and enabled, here’s exactly what happens on a live call:

1. Each node is a phase. Only the current node’s prompt and tools are active, which keeps the agent on-script.
2. The opening node greets (its First Message), then the caller speaks.
3. After each caller turn the Director extracts the current node’s declared variables from what the caller said.
4. The engine resolves the transition in a fixed order: equation edges first (deterministic, first-true-wins) → then the Director’s chosen prompt-edge (only if the node’s require[] variables are satisfied) → then global enter-conditions → else stay on the node and re-ask.
5. Route, then reply. The flow lands on the destination node first, and that destination node produces the reply.
6. Internal nodes pass through. Condition, Set Variable, and Transfer nodes run instantly without taking a caller turn, then hand off to the next conversational node, which speaks.
7. API errors don’t hang the call. On a failed API Request, the node follows its error edge so the flow always moves on.
8. Globals can interrupt and return. A Global node becomes reachable from any interruptible node when its enter-condition matches (e.g. the caller raises a pricing objection or asks for a human). It handles the situation, then returns the caller to the node they were on. A per-global cooldown (in turns) stops it from re-triggering immediately.

For the full transition mechanics, edge types, and operator reference, see Transitions, Variables & the Director.

​

Alpha limitations

​

Next steps