/
/
Voice-Agent Tasks (𝜏³)

Available on Request · Sample Available

Voice-Agent Tasks (𝜏³)

Full-duplex voice-agent tasks.

Volume

0

Volume

Scalable Volume

Coverage

Agentic Workflows

0.0 %

Gemini 3.1 Flash Live

Targets

Tau3-Knowledge

Modality & Delivery

AudioText→Text

Formats

JSON

What You Receive

Full-duplex voice-agent tasks testing banking use-cases across customers. Includes call recordings, timestamped transcripts, task definitions, gold actions, a customer database, prompts, and policy documents.

Coverage

Agents must identify the correct account, search bank policies, verify customer claims, and take the appropriate action. Tasks are scored against gold actions and database outcomes with no human or LLM judge.

Package Structure

  • audio/
    • task_1.wav
    • task_2.wav
    • task_3.wav
  • data/
    • db.json
    • documents/Knowledge-base Documents
      • doc_bank_accounts_bank_accounts_(general)_001.json
      • doc_credit_cards_virtual_card_management_001.json
      • doc_savings_accounts_gold_account_001.json
    • prompts/
      • components/
        • additional_instructions.md
        • policy_header.md
        • shell_instructions.md
      • agentic_search.md
      • agentic_search_write.md
      • all_tools.md
      • classic_rag.md
      • full_kb.md
      • grep_only.md
      • no_knowledge.md
      • required_docs.md
    • tasks/
      • task_1.json
      • task_2.json
      • task_3.json
      • task_4.json
    • split_tasks.json
  • transcripts/
    • task_1.txt
    • task_2.txt
    • task_3.txt
    • task_4.txt

Differentiators

  • Full-duplex voice tasks include recordings, timestamped transcripts, gold actions, database state, and policy documents.

Record Schema

Results

ModelMetricResult
Gemini 3.1 Flash LivePass@125%

Methodology & Quality

Methodology

SME involvement

Domain-expert curated with up to L2 review

Stressors

Fraud triage across customers with multiple similar accounts and cards under real-time voice interaction.

Design criteria

  1. Grade against gold actions and database outcomes without a human or LLM judge.

Build assumptions

Quality approach

  • Deterministic grading against gold actions and database outcomes.

Known limitations

Customization & Scalability

Adoptable to custom tooling and environments. Additional volume in progress