Aug 29, 2026

AgentVox

Link to open source: https://github.com/NickyY28/AgentVox

AgentVox: Adaptive Multi-Agent Voice AI Interview Platform

AgentVox is an adaptive, real-time voice AI interview platform designed to make AI-powered interviews more conversational, evidence-driven, and transparent. Built for the EchoSphere: Agora Conversational AI Hackathon 2026, AgentVox uses real-time voice interaction through Agora combined with a multi-agent AI architecture to conduct interviews that adapt to what a candidate actually says.

Traditional AI interview systems often follow predefined question lists. Every candidate may receive essentially the same sequence regardless of their resume, experience, role, or previous answers. This creates a major limitation: a good interview is not simply a collection of questions and answers. A good interviewer listens, identifies claims, asks for evidence, explores reasoning, and decides what should be asked next based on the conversation.

AgentVox is designed around this principle.

The Problem

Current AI interview experiences can suffer from several issues:

Generic questions: Candidates may receive fixed questions without meaningful adaptation to their background or the job requirements.
Robotic interactions: Systems can behave more like question generators than genuine conversational interviewers.
Irrelevant evaluation signals: Accent, appearance, eye movement, gestures, or vocal characteristics can become proxies for assessment instead of focusing on demonstrated competency.
Low transparency: Candidates and recruiters may receive scores without understanding the evidence behind those evaluations.

The central problem AgentVox addresses is therefore not simply how to automate interviews, but how to make AI evaluation more relevant, adaptive, and trustworthy.

The AgentVox Approach

AgentVox begins by understanding the candidate's resume and job description. From this information, it identifies relevant competencies and creates an interview strategy.

During the interview, the candidate communicates naturally through voice. Instead of following a rigid question sequence, AgentVox continuously considers the conversation and determines what should happen next.

For example:

Candidate: “I reduced our API latency by 40% in my last role.”

Instead of simply moving to the next predefined question, AgentVox can investigate the claim:

AgentVox: “How did you measure the before and after latency?”

The candidate might explain that they used Datadog to track p95 response times. AgentVox can then continue:

AgentVox: “What was the baseline, and which specific change contributed most to that reduction?”

This creates a progressively deeper interview:

Claim → Evidence → Deeper Probe → Reasoning → Evaluation

The next question is influenced by the candidate's previous answer rather than being predetermined.

Multi-Agent Intelligence

AgentVox is designed around specialized AI responsibilities rather than relying on a single generic agent.

The architecture can include specialized roles such as:

Interviewer Agent — manages the live interview conversation.
Resume/Context Agent — understands the candidate's resume, role, and job requirements.
Evidence/Follow-up Agent — identifies claims and determines where additional evidence or clarification is needed.
Evaluation Agent — evaluates demonstrated competency based on the evidence collected.
Final Assessment Agent — organizes the interview findings into a structured evaluation.

These agents work together through the interview process, allowing AgentVox to maintain context while exploring competency gaps and candidate reasoning.

Real-Time Voice with Agora

Agora is a core part of AgentVox's conversational experience.

The candidate communicates with AgentVox through real-time voice, allowing the interview to feel closer to an actual conversation rather than a text-based questionnaire.

The conceptual flow is:

Candidate
    ↓
Agora Real-Time Voice
    ↓
AgentVox Conversation Layer
    ↓
Multi-Agent Intelligence
    ↓
Context + Evidence + Evaluation
    ↓
Adaptive Next Question
    ↓
Agora
    ↓
Candidate

This real-time interaction is important because AgentVox's intelligence is designed to operate during the conversation, rather than only analyzing the interview after it has ended.

Evidence-Based Evaluation

A key principle of AgentVox is that candidates should be evaluated on job-relevant evidence.

Instead of relying heavily on superficial behavioral signals, the system focuses on what the candidate can demonstrate through their answers, explanations, decisions, and reasoning.

AgentVox can track:

Which competency is being evaluated
What evidence the candidate provided
Which claims have been substantiated
Where additional evidence is missing
How confidently the system can make an assessment

This creates an evidence trail that can make the evaluation easier to understand and review.

Uncertainty and Human Review

AgentVox also follows an important principle:

When AI isn't sure, it shouldn't pretend.

If the available evidence is insufficient to confidently evaluate a competency, AgentVox can flag the evaluation for human review rather than presenting an uncertain judgment as an unquestionable score.

This introduces a human-in-the-loop approach where AI assists the hiring process while acknowledging situations where additional human judgment may be appropriate.