Selection & runtime

You don't route procedures manually. The agent selects the right one from the conversation, then runs through its steps. This page explains how that works.

Telling the agent when to use a procedure

Each procedure has selection criteria you set when authoring it:

• When to use — a plain-English description of the situation, like "the customer wants to cancel their subscription." This is the main signal.
• Trigger examples — "use when" and "avoid when" phrases that sharpen the choice.
• Rulesets — optional structured conditions that pre-filter candidates cheaply before the agent decides.

How selection happens

Each turn, the agent:

1. Pre-filters candidates using any rulesets (fast, deterministic).
2. Picks the best match among the survivors based on the conversation and each procedure's "when to use" and examples.
3. Pins the version — the selected procedure's active version is locked in for that conversation.

If nothing matches, the agent stays in free-form mode and just uses its persona and tools.

During the conversation

While running a procedure, two things can interrupt the flow:

• Digression — the customer asks something off-script. The agent handles it and then returns to where it left off.
• Handoff — the situation needs a person. The agent escalates to a human and leaves the procedure.

Version pinning

Once a conversation has started running a procedure, it keeps the pinned version even if you publish a change mid-conversation. The next conversation that selects the procedure gets the new active version. This means edits never disrupt a conversation that's already in flight. See Procedure versions.

Per-agent overrides

The same procedure can behave differently on different agents. When you attach a procedure to an agent, you can:

• Override its when to use, examples, and rulesets for that agent.
• Enable or disable the attachment.
• Set a priority so it's considered before or after others.

Next steps