Multi-Agent System for Interview Analysis and Report Generation at Scale

listenlabs · Tech · 2026 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: User research analysis

ListenLabs needed to conduct and analyze hundreds to thousands of interviews, surveys, and focus groups, turning qualitative data into reports, charts, clips, and presentations.

Constraints:

The platform covers the whole lifecycle from study design through collection and reporting.
Interview agents must support voice, screen sharing, images, and audio/video conversations.
Research analysis can take around 30 minutes for comprehensive reports.
Quality requires evaluation agents, trace analysis, contextual prompts, and sandboxed code execution.

Architecture

ListenLabs built three primary agents plus subagents: Composer creates discussion guides through an artifact interface, an interview agent conducts multimodal voice interviews, and a research agent processes large volumes of qualitative data to produce insights, charts, clips, and PowerPoint decks. Parallel subagents process interviews, evaluation agents check quality, and trace analysis drives improvement.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  composer-agent["Composer Agent"]
  interview-agent["Interview Agent"]
  research-agent["Research Agent"]
  parallel-subagents["Parallel Subagents"]
  evaluation-agents["Evaluation Agents"]
  sandboxed-code-execution["Sandboxed Code Execution"]
  composer-agent --> interview-agent
  interview-agent --> research-agent
  research-agent --> parallel-subagents
  research-agent --> evaluation-agents
  research-agent --> sandboxed-code-execution
  parallel-subagents --> research-agent
  evaluation-agents --> research-agent
  sandboxed-code-execution --> research-agent

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Composer Agent`	Collaboratively builds discussion guides through an artifact-based interface.	`Interview Agent`
`Interview Agent`	Conducts voice-based multimodal interviews with participants.	`Research Agent`
`Research Agent`	Analyzes qualitative data and generates insights, charts, clips, and reports.	`Parallel Subagents`, `Evaluation Agents`, `Sandboxed Code Execution`
`Parallel Subagents`	Process many interviews simultaneously.	`Research Agent`
`Evaluation Agents`	Perform quality control over generated research outputs.	`Research Agent`
`Sandboxed Code Execution`	Runs code for analysis artifacts such as charts and reports.	`Research Agent`

Data flow

A user collaborates with Composer to create a discussion guide. The interview agent runs voice-based multimodal sessions with participants. The research agent processes completed interviews in parallel, uses sandboxed code execution and subagents to create analyses, charts, clips, and presentation artifacts, and evaluation agents check quality.

“The platform handles the complete lifecycle from study design through data collection to automated analysis and reporting.”summary, stated

“parallelized sub-agent execution for processing hundreds of interviews simultaneously”summary, stated

Why this architecture

ListenLabs uses separate agents because study design, live interviewing, and large-scale research analysis have different interaction patterns and tool needs. The artifact interface avoids full-document rewrites, while parallel research subagents make large interview sets tractable.

“ListenLabs operates three primary agents that handle distinct phases of the research process, along with numerous specialized sub-agents.”summary, stated

“earlier versions where the AI would completely rewrite documents, which proved problematic for longer documents”summary, stated

Models used: unspecified multimodal and analysis LLM agents

Scale

The platform processes hundreds to thousands of interviews, surveys, and focus groups.
“processes hundreds to thousands of user interviews, surveys, and focus group feedback.”summary, stated
The system evolved over about two years.
“The system has undergone multiple architectural iterations over approximately two years”summary, stated
Comprehensive reports can take around 30 minutes.
“a long-running asynchronous analysis that can take around 30 minutes for comprehensive reports”summary, stated

Alternatives considered (roads not taken)

Use basic RAG for qualitative research

Rejected because: The system evolved from basic RAG to a multi-agent architecture as requirements matured.

“evolving from basic retrieval-augmented generation implementations to a complex multi-agent architecture.”summary, stated

Let the AI rewrite complete discussion guides

Rejected because: Full rewrites were problematic for long documents where users wanted targeted edits.

“users often want targeted changes rather than complete regeneration.”summary, stated

Patterns

`multi-agent-orchestration` (architecture)

Composer, interview, research, subagent, and evaluation agents cover the research lifecycle.

“a complex architecture featuring three primary agents: a study creation agent (Composer)”summary, stated

`batch-offline-processing` (execution)

Research analysis runs asynchronously and parallelizes subagents across many interviews.

“parallelized sub-agent execution for processing hundreds of interviews simultaneously”summary, stated

`eval-driven-development` (quality)

Custom evaluation agents and trace analysis are used for quality control and continuous improvement.

“custom evaluation agents for quality control, contextual prompt engineering, code execution in sandboxes, and sophisticated trace analysis”summary, stated

Tradeoffs

Use artifact-based study creation

Chose Editable discussion-guide artifact with AI and human operations over Chat-only full-document regeneration.

Rationale: The artifact pattern supports targeted edits and undo/change tracking.

“allowing for undo functionality and change tracking.”summary, stated

Run long asynchronous research analysis

Chose Comprehensive reports that can take around 30 minutes over Only fast synchronous Q&A over interview data.

Rationale: Deep reporting requires charts, clips, summaries, and PowerPoint outputs across many interviews.

“creating charts, summarizing findings, cutting video clips, and even generating PowerPoint presentations”summary, stated

Gotchas (what bit them)

AI document editing can be too destructive if it rewrites whole artifacts.

Root cause: Users often want targeted discussion-guide changes, not complete regeneration.

Resolution: Composer uses an artifact interface with unified edit history and direct/AI edits.

“The system maintains a unified edit history where both human and AI changes are tracked as edit operations”summary, stated

Interview agents should not over-interrupt participants.

Root cause: The research goal is to let participants express themselves fully, including natural pauses or rambling.

Resolution: The interview agent deliberately emphasizes listening rather than jumping in.

“The platform emphasizes listening rather than interrupting, deliberately avoiding jumping in during natural pauses in conversation”summary, stated

Evaluation

ListenLabs uses custom evaluation agents, trace analysis, and prompt iteration for quality control.

Metrics: hundreds to thousands of interviews · approximately two years of iterations · around 30 minutes for comprehensive reports

“custom evaluation agents for quality control, contextual prompt engineering, code execution in sandboxes, and sophisticated trace analysis for continuous improvement.”summary, stated

Outcomes

ListenLabs covers study design, data collection, analysis, and generated reporting.
The platform can process large interview sets with parallel subagents.
The architecture matured from basic RAG to a multi-agent lifecycle system.

The case cites hundreds to thousands of interviews, two years of iteration, and 30-minute comprehensive reports.