New Framework Automates Software Requirements Interviews With Dual-Agent LLM System

Gathering detailed and usable software requirements directly from end-users is a foundational step in building effective applications, but the traditional approach—manual laddering interviews—remains labor-intensive and difficult to scale. Researchers have now developed LadderTeam, an automated framework that uses a pair of AI agents to conduct these interviews at scale while maintaining quality.

What the Framework Does#

LadderTeam employs a dual-agent architecture built on large language models. An interviewer agent drives the conversation using one of three distinct questioning strategies—Attributes, Consequences, Values (ACV); 5-Whys analysis; or Jobs To Be Done (JTBD)—to extract granular, actionable requirements from user feedback about wireframes. Simultaneously, a Judge agent monitors the exchange in real time, scoring probe-response pairs and enforcing guardrails to prevent the conversation from drifting off-topic.

Novel Testing Methodology#

Rather than relying on human participants (which introduces variable confounds), the researchers tested LadderTeam using a controlled simulation approach. They created scripted ground-truth interview transcripts representing different personality types and evaluated how well the system's probes elicited actionable insights. This isolated probe quality as the sole experimental variable.

Results#

Across 216 simulated interviews, LadderTeam achieved 99.1% chain convergence and matched ground-truth actionable responses 81.0% of the time (86.1% for reluctant personalities, 75.9% for terse personalities). Critically, the system maintained zero topic drift across all runs, suggesting the Judge agent's guardrails functioned effectively.

Open Science and Future Availability#

The team plans to open-source all evaluation code, transcripts, inputs, and a live demonstration platform following publication, enabling the research community to build on and validate the work.