Pathways

Every ClearTalk agent is driven one of two ways. On the agent, an Input Type toggle lets you choose:

• Prompt — plain-English instructions (the agent's Agent Instructions). You describe the job, and the agent works out what to say in the moment.
• Pathway — you draw the conversation as a map of connected steps, and the agent follows it.

Both are first-class. Neither is the "advanced" option. The right answer depends on how much variation you're willing to accept in exchange for how natural the conversation sounds.

Start prompt-based when…

• The call has one job — book a viewing, qualify a lead, confirm an appointment.
• You want it to sound conversational and adaptive, handling questions in whatever order they come.
• The consequences of the agent phrasing something differently each time are low.
• You're still learning what callers actually say. Instructions are much faster to iterate on.

Most agents should start here. You can always build a pathway later, and the instructions you wrote become the raw material for it.

Build a pathway when…

• The same things must happen every time — the same questions asked, the same disclosures given, in the same order.
• The conversation genuinely branches and each branch does something different: new customer versus existing, in-network versus out, interested versus not.
• You need to do something mid-call — look up an order, check a calendar, send a text, capture a document, transfer to the right department.
• You need to test it. Only pathways can be tested with simulated callers before you point real leads at them (see Testing your pathway).
• More than one person maintains it. A map is reviewable; a wall of instructions is not.
• Compliance is involved. When a script is regulated, "the agent usually says this" isn't good enough.

Every pathway has two sides

Worth knowing before you build anything, because it changes how freely you can work.

A pathway isn't one thing — it's a working copy you edit and a live version that answers real calls. They're separate. You can rebuild a pathway from scratch over a weekend, save it thirty times, and Monday's callers still hear exactly what they heard on Friday. Nothing you do reaches a customer until you deliberately publish it.

That's why a pathway is the safer choice for anything you'll be changing regularly with real traffic running: there is no moment where a half-finished edit is live. Prompt-based agents have no equivalent — an edit to the instructions is the edit.

Publishing, versions, and rolling back are covered in Publishing and versions. For now just carry the idea with you: saving is always safe.

The trade-off, honestly

A pathway is not automatically better. A badly built one — long chains of rigid statements with no room to breathe — sounds worse than good instructions. The goal is structure where structure matters, and freedom everywhere else. Individual steps can still generate their reply from a prompt, so you can pin down the parts that matter and let the agent be natural in between.

Moving from a prompt to a pathway

You don't have to start over. Take the instructions you already have, and:

1. List the moments that must happen in order — that's your spine.
2. List the questions callers actually ask (your transcripts will tell you) — those become branches.
3. Build the spine first and publish it. Add branches as real calls reveal them.

Where to go next