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.
Customer Support / T2-M4
What this task measures
The caller is on the line with Kestrel Electronics for one locked request: Washington drone restocking quote. They keep repeating that ask through a callback offer, a different ticket, or an unrequested fix. We included it as a medium returns and refunds case because window and restocking math has to be spoken before a return is opened; state-law exclusions are the usual miss. The agent succeeds only if it stays on that request, reaches the outcome by following the verification then returns 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 $0.122
0:00 / 1:13 · audio · trace 813281 · click a span to play, drag to scrub, scroll to pan
Audio
Speech Audio
Agent Audio
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 handoff path did not match. Combined cannot pass unless every gate does.
Expected versus actual tool calls on this conversation.
Expected
Actual
Parameters
None recorded.
None recorded.
Output
None recorded.
None recorded.
Parameters
None recorded.
None recorded.
Output
None recorded.
None recorded.
Parameters
- card_last4: 7735
- postal_code: 98042
Output
None recorded.
None recorded.
Parameters
None recorded.
None recorded.
Output
None recorded.
None recorded.
Parameters
None recorded.
None recorded.
Output
None recorded.
None recorded.
Parameters
- opened: true
- order_number: KE-4492551
Output
None recorded.
None recorded.
Parameters
+ topic: returns
Output
+ ok: true
Parameters
+ topic: restocking
Output
+ ok: true
Parameters
+ fee: restocking fee
Output
+ ok: true
Parameters
+ topic: marketplace
Output
+ ok: true
Open a row for parameters and outputs. Required tools must appear. Extra industry tools are allowed. Missing: transfer_to_verification, identify_customer, verify_identity, get_customer_summary, transfer_to_returns, check_return_eligibility. Extra: get_policy, get_policy, get_fee, get_policy.
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 through the shared prefix, then red from the first wrong hop.
Expected path
transfer_to_verification → transfer_to_returns
Actual path
stay on reception
The actual hangup matches the expected hangup.
Locked caller script. This is the technical definition. Copy it as written.
Open with exactly: "I want to know what it would cost to return the Skyward Vireo 3 drone I bought in Washington. The box is opened." Give identity when asked. Order KE-4492551. You only want eligibility and the restocking fee. Do not start the return. Hang up once they say 15 percent, one hundred forty-nine ninety-nine. Once they have finished what you asked and confirmed it, say exactly: "Thank you, that's all I needed." and end the call. Do not start a second request.
