Listen CTO explains self-reviewing agents for qualitative research
Florian Juengermann, co-founder and CTO of Listen, describes how the startup's multi-agent system analyzes hundreds of interviews and surveys at scale, including a self-reviewing feedback subagent that catches errors during long async runs.
Score breakdown
The virtual table architecture and self-reviewing subagent pattern offer concrete, replicable design ideas for agent engineers building systems that must process large volumes of unstructured data with quality guarantees.
- 01Listen's platform includes three agents: a Composer (study creation), an interviewing agent, and a Research Agent for report generation.
- 02The Research Agent uses a 'virtual table' architecture where each row is a response and each column is an extracted feature or question.
- 03Small models run map-reduce classification across thousands of open-ended responses.
Florian Juengermann, co-founder and CTO of Listen, appeared on LangChain's "Max Agency" podcast to detail how Listen evolved from a simple RAG bot into a multi-agent system capable of analyzing qualitative research at scale. The platform's core Research Agent uses a "virtual table" architecture: every row represents a single interview response, and every column represents a question or extracted feature. The agent creates new columns dynamically to surface structured signal from unstructured, media-rich conversations. Small models perform map-reduce classification across thousands of open-ended responses, enabling massive parallelization — including a hardcoded workflow that spawns 500 agents at once and aggregates their outputs in a specific way.
Juengermann also discusses contextual prompt engineering as an alternative to skills-based approaches, including a feature for emotional understanding derived from video and audio — not just text.
Listen's platform comprises three distinct agents: the Composer, an interactive creation agent that works collaboratively with users on a discussion guide artifact; an interviewing agent that conducts conversations with hundreds or thousands of users in parallel; and the Research Agent that produces the final report. Quality control is handled by a dedicated reviewer subagent that "knows what a good report looks like," used both as an error-catcher during long async runs and as an evaluation system in the live runner. Juengermann also discusses contextual prompt engineering as an alternative to skills-based approaches, including a feature for emotional understanding derived from video and audio — not just text. The PowerPoint subagent was completely rebuilt using Claude's Code SDK, and the conversation also covers when to trigger long-running agents versus surfacing early results.
Key facts
- 01Listen's platform includes three agents: a Composer (study creation), an interviewing agent, and a Research Agent for report generation.
- 02The Research Agent uses a 'virtual table' architecture where each row is a response and each column is an extracted feature or question.
- 03Small models run map-reduce classification across thousands of open-ended responses.
- 04A hardcoded workflow can spawn 500 agents simultaneously and aggregate their outputs.
- 05A self-reviewing subagent monitors report quality during async runs and doubles as an evaluation system in the live runner.
- 06The PowerPoint subagent was completely rebuilt using Claude's Code SDK.