Working paper / 2026
Multi-Industry Voice Agent Simulation Bench
Abstract
Multi-Industry Voice Agent Simulation Bench (MIVAS) is an indicator of voice AI performance across economic sectors. MIVAS is comprised of tasks and verifiers with granular rewards executed in production-grade multi-agent RL environments across industries where voice AI is being adopted the fastest. Each task places a model harness inside a multi-agent graph with production-length prompts, tools, and isolated state. Deterministic verifiers score tool use, handoffs, and final database state; pass@1 and passk measure single-run capability and repeated-run reliability.
Clone the repository, pair a model harness with an industry task suite, and reproduce the evaluation environment.
Healthcare / C2-H1
What this task measures
The caller is on the line with Straus Dermatology for one locked request: Cancel tomorrow's follow-up. They keep repeating that ask through a waitlist offer, a callback, or a different errand. We included it as a hard appointment management case because calendar writes are a core front-desk job, and they fail in characteristic ways: the wrong visit moves, identity is skipped, or a cancel becomes a reschedule. The agent succeeds only if it stays on that request, reaches the outcome by following the identity then scheduling path, and leaves the record exactly as this case requires: the right row written or cleared, and nothing extra.
One model harness at a time, with a representative pass and fail conversation when both exist. OpenAI Realtime 2.1 is the default view.
Recording
Conversation cost $0.268
0:00 / 2:08 · audio · trace 801784 · click a span to play, drag to scrub, scroll to pan
Audio
Speech Audio
Agent Audio
Handoffs
Tool Calls
Errors
Transcript
How a conversation passes
Combined pass is the logical and of three deterministic gates. Tools, handoff path, and final database state must all pass on the same conversation. One miss fails the call.
Combined
Fail
This sample fails because tools and final database state did not match. Combined cannot pass unless every gate does.
Expected versus actual tool calls on this conversation.
Expected
Actual
Parameters
+ from: reception
- next_intent: scheduling
+ to: identity
Output
None recorded.
None recorded.
Parameters
+ from: identity
+ to: scheduling
Output
None recorded.
None recorded.
Parameters
dob: 1990-04-12
dob: 1990-04-12
full_name: Jordan Lee
full_name: Jordan Lee
Output
- data: {"verified":true}ok: true
ok: true
Parameters
None recorded.
None recorded.
Output
ok: true
ok: true
Parameters
+ is_new_patient: false
+ urgency: routine
+ visit_class: medical
Output
+ ok: true
Parameters
+ location_id: loc_park_ave
Output
+ ok: true
Parameters
appointment_id: 1
appointment_id: 1
cancellation_reason_code: patient_request
cancellation_reason_code: patient_request
+ fee_disclosed_and_accepted: false
Output
- data: {"status":"fee_disclosure_required"}ok: true
ok: true
Parameters
- appointment_id: 1
- cancellation_reason_code: patient_request
- fee_disclosed_and_accepted: true
Output
- data: {"status":"cancelled"}- ok: true
Parameters
- appointment_type_code: MED_FOLLOWUP
- earliest: 2026-08-24
- location_ids: ["loc_park_ave"]
Output
- data: {"status":"added"}- ok: true
Open a row for parameters and outputs. Required tools must appear. Extra industry tools are allowed. Missing: cancel_appointment, join_waitlist. Extra: classify_visit_request, list_locations.
Full industry architecture, same graph as the industry page. The left diagram marks the expected specialist path in green. The right diagram marks the path this conversation actually took in green, because it matched.
Expected path
transfer_to_identity → transfer_to_scheduling
Actual path
transfer_to_identity → transfer_to_scheduling
Rows that differ between the expected hangup and the actual hangup.
appointments · 1
Expected
Actual
- status: cancelled
+ status: booked
missing · waitlist · 1
Expected
Actual
- appointment_type_code: MED_FOLLOWUP
- earliest: 2026-08-24
- id: 1
- latest: 2026-09-30
- patient_id: pat_jordan_lee
Locked caller script. This is the technical definition. Copy it as written.
You are an existing Straus patient. You have a follow-up on August 20, 2026 at Park Avenue and you need it cancelled outright because you are travelling. Answer their identity questions once with your name and date of birth, confirm if they read it back, then go straight to the cancellation. You will not rebook. If they offer another time, refuse it and say you only want this visit cancelled. After they explain a missed-visit fee, say you accept it and still want the August 20 visit cancelled. After it is cancelled, if they offer the waitlist for a later opening, say yes and ask to be listed from August twenty-fourth through September thirtieth at Park Avenue — those calendar dates only. Do not name a clock time, and if they ask for a latest time of day say end of September thirtieth is fine. Then thank them and end the call. Do not ask to book or reschedule a new visit. Do not ask to add insurance, for a callback, to speak to a person, a bill, a payment, the patient portal, or a text. Do not ask for the street, floor, suite, or office id. Do not volunteer a ZIP. Give the waitlist window once. If they ask again, say you already gave August twenty-fourth through September thirtieth — do not agree to a second listing. Do not ask them to text you an appointment confirmation.