Build the agent, hear it, ship it

The same agent answers an inbound line, runs an outbound campaign, sits behind a branch of a phone menu, and staffs a chat widget on your site. You build it once. Where it gets used, and what data it is handed when it starts, is a separate decision.

That is why the wizard asks about the agent itself and nothing about the campaign. Retry policy, calling window, which number it answers on, which contact list it runs against: none of that lives on the agent, so changing any of it never means rebuilding one.

Navigation runs down the left of the dialog rather than across the top. That gives the content the full width and leaves each step room to say what it currently holds, so you can see the whole configuration without opening every step.

Vocalis Realtime is the default and the one most lines should run. It is our speech-native stack, managed end to end, tuned for sub-second responses and natural turn-taking on phone audio.

The other two are escape hatches for teams with a specific reason to assemble their own. A custom pipeline lets you choose transcription, reasoning and speech separately, from six vendors. Speech to speech runs a third-party realtime model with its own built-in voice and no transcription hop. The catalogue behind both is on the models page.

A field is anything the agent should say differently from one call to the next. Fields can be filled from a contact list column, from the API request that started the call, or by a step earlier in a workflow. A field can also be static, which is how a plan name or a department gets in.

Write {{claim_number}} in the prompt and the fields step notices, offers to add it, and the prompt stops being a promise the agent cannot keep. Connect a contact list and its columns are offered the same way, so the fields you add always match data that exists.

Tools are what turns an agent from something that answers into something that acts. Point it at an OpenAPI document and each endpoint arrives as a callable tool, with the description the agent uses to decide when to reach for it. A claims lookup, an eligibility check, a callback booking, a ticket.

Transfers are tools too. Warm and cold transfer take a set of named destinations rather than a number, so the prompt can say who to reach instead of hardcoding a phone number that will change. Knowledge base queries are tools as well, which is what lets an agent decide it needs to look something up rather than answering from memory.

Tools can be tested on their own before an agent ever calls one, and some are gated behind an integration so they cannot be switched on before the thing behind them is connected.

Attach a knowledge base and the agent answers from your plan documents instead of from what it happens to know. Upload the evidence of coverage, the formulary, the prior authorization policy. The content is indexed by section, and the answer comes back with the section it was read from.

Grounding is the difference between an agent that is plausible and one that is correct. It is also what makes an answer checkable after the fact: someone can go and read what the agent read.

The composer inserts fields and tool names rather than making you remember them, and it warns on the things that break calls quietly: a field that does not resolve, a transfer target with no description for the agent to route on, a knowledge base referenced but not attached.

Underneath whatever you write, every call carries a fixed set of safety instructions that a prompt cannot override. They are what keeps a caller from talking the agent out of its own rules, and they are not editable, on purpose.

Every step has a test bar. Press it and you are on a call with the agent exactly as configured, in the browser, with the fields filled in. Change a line of the prompt, press it again. It is the fastest way to find out that an instruction that reads well is unusable out loud.

Every save is a version, attributed and timestamped, with a reason if you gave one. The history panel shows what changed between any two: prompt text, but also the model, the voice, the fields and the tools, since those break a call just as easily as wording does.

Rolling back restores an earlier version as the live one without deleting anything that came after it, so a bad afternoon costs you a click rather than a rebuild.

Once an agent exists it can be pointed at work in four ways, and none of them change the agent.

Campaign settings live with the campaign, which is why the same agent can run inside quiet hours in one and a wider window in another. The window is evaluated in the contact’s local time, and retries back off rather than redialling immediately.

An agent that works can be saved as a template, so the next line starts from something proven instead of a blank prompt. Agents can also be shared to other workspaces directly, which keeps one canonical version instead of five copies drifting apart.

Who can edit a shared agent, and who can only run it, is decided by role. That is covered on the security page.