← Back to harness
Pillar 3 · Supervisor architecture

A great talker that follows the process.

The supervisor sits around the live speech-to-speech model. It keeps the call on the right stage, enforces guardrails, decides when tools may run, and handles barge-in and timeouts.

OutcomeThe agent sounds natural and still follows the customer's process, every step, every call.
The idea

The model talks. The supervisor runs the call.

A raw speech-to-speech model is very good at conversation. It is not built to follow a business process for twenty minutes. Dinodial wraps every live call in one supervisor that holds the process, so the model can focus on the talking.

CallerPhone or browser widget↔
Supervisor · one per call
StageKnows which step of the flow the call is on, and where each outcome routes next.
GuardrailsEvery tool call is checked against the current stage before it can run.
Speech-to-speech modelGemini Live
Turn controlDecides when caller speech cuts the agent off, and what to do with a silent caller.
WatchdogsDeadlines on every live call, so a stuck call cannot hold capacity forever.
Tools and outcomeCustomer systems, call result↔

The supervisor reads one event at a time from the caller, the model, tool results and timers. Its decisions are plain data, so the same logic is unit tested and reused for phone calls, the widget and simulation.

1 · Stages

The process lives in the flow, not in a long prompt.

Each stage has its own instruction, the tools it may use, a barge-in setting and its outcomes. This is a real health insurance intake agent: eleven stages, each routing to the next, with wrap-up and premature close handled as separate post steps.

Flow graph of an insurance intake agent with eleven stages123
  1. 1Start stage. The call always opens here, after the greeting clip.
  2. 2One section per stage. Family history, alcohol, tobacco. The model only sees the job for the stage it is in.
  3. 3Post steps. Wrap-up records the outcome. Premature close is its own path, so an early hang-up is labelled, not lost.
Stage editor for the members stage12345
2 · Inside one stage

Stage 2 of 11: who is covered.

Business rules sit next to the step they belong to. Here, only relatives with an insurable interest can be added.

  1. 1Barge-in. Set per stage. Off means a caller cannot cut a required disclosure short.
  2. 2Allowed tools. Only save_section may run in this stage.
  3. 3Instruction. What to say, what to ask, and how to handle edge cases, scoped to this step.
  4. 4Routing. Success moves the call to ages. Failure has its own route.
  5. 5Test and publish. Test against the draft on gemini-prod before anything goes live.
Tool guardrails: repeat limits, in-flight policy and scope1234
3 · Guardrails

The model can ask. The supervisor decides.

Every tool call passes five checks in order: is it declared, is it allowed in this stage, is one already in flight, has it hit its repeat limit, then dispatch.

  1. 1Repeat limit. record_outcome runs once. save_section at most five times.
  2. 2In flight: reject. A second call while the first is still running is refused.
  3. 3Sync. The model waits for the result before it continues.
  4. 4Scope. Each tool writes only where it is allowed to.

A rejection never names the stage that would allow the tool, so the model cannot learn to skip a step. Results that arrive after the call has moved on are dropped as stale.

4 · Barge-in and watchdogs

Turn taking and timeouts are rules, not hopes.

Barge-in

One path for every vendor.

When the caller talks over the agent, the vendor reports it and one controller decides whether to clear the audio. A protected reply is not cleared. Lens shows the discarded audio in red.

Silent callers

Prompt, then end cleanly.

After 15 seconds of silence the agent plays an idle prompt. After another 15 seconds with no answer, the call ends as unresponsive instead of hanging open.

Stuck calls

Every call holds a lease with a deadline.

If audio never connects (120 seconds on phone, 30 on the widget) or a call runs past its hard limit plus 60 seconds, capacity is freed first and cleanup follows.

Dead ends

Misses escalate.

If the model keeps reporting an outcome the stage does not know, the flow counts misses and, after three by default, routes the call to failure.

What this means for a deployment

Gemini Live does the conversation. Dinodial makes it follow the process.

01

The script holds for the whole call.

Steps, disclosures and routing live in the flow. A long call does not drift off the process because the model forgot part of a prompt.

02

Actions happen at the right moment.

A booking, a save or a close can only run in the stage that allows it, a limited number of times, one at a time.

03

Calls end on purpose.

Silent callers, early hang-ups and stuck calls each end with a labelled outcome, so reports match what really happened.

04

Fixes are small and safe.

A problem in one step is fixed in that stage, tested against the draft and published as a new version. Nothing else moves.