AI agents that build their own data tables for analysis
ListenLabs AI Co-Founder and CTO Florian Juengermann describes an agent system that analyzes data by dynamically creating table columns — such as "user sentiment" — and automatically filling in values for each row using constrained sub-agents.
Score breakdown
The map-reduce-style sub-agent pattern for dynamic column generation offers a concrete architectural blueprint for building structured, scalable data-analysis agents.
- 01Florian Juengermann is Co-Founder and CTO of ListenLabs AI.
- 02The main agent structures data as a table: each row is a response, each column is a question or extracted feature.
- 03The agent can dynamically create new columns — e.g., 'user sentiment toward a specific topic' — and auto-fill values for every row.
In this episode of Max Agency, LangChain CEO Harrison Chase interviews Florian Juengermann, Co-Founder and CTO of ListenLabs AI, about the architecture behind his data-analysis agents. Rather than working directly with raw file structures, the main agent organizes data as a table where each row represents a response and each column represents a question or extracted feature. The agent is capable of creating new columns on the fly — for instance, generating a "user sentiment toward a specific topic" column that can hold either a free-text summary or a categorical value such as positive, negative, or neutral.
Once the table is populated, tools such as Python can be used to chart and visualize the results, enabling what he describes as AI-powered data analysis on autopilot.
To fill in the values for each row in a newly created column, the system launches what Juengermann calls sub-agents, though he emphasizes these are highly constrained: they have no decision-making latitude and simply execute a single classification call. He likens the overall pattern to a map-reduce operation. Once the table is populated, tools such as Python can be used to chart and visualize the results, enabling what he describes as AI-powered data analysis on autopilot.
Max Agency is a podcast focused on how real-world AI agents are designed, deployed, and iterated on, covering topics from architecture decisions and evaluations to tooling and failure modes.
Key facts
- 01Florian Juengermann is Co-Founder and CTO of ListenLabs AI.
- 02The main agent structures data as a table: each row is a response, each column is a question or extracted feature.
- 03The agent can dynamically create new columns — e.g., 'user sentiment toward a specific topic' — and auto-fill values for every row.
- 04Column values can be open-ended summaries or categorical labels such as positive, negative, or neutral.
- 05Row-filling is handled by highly constrained sub-agents that perform a single classification call, described as analogous to a map-reduce operation.
- 06The populated table can be charted using tools like Python.
- 07Max Agency is hosted by Harrison Chase, CEO of LangChain, and focuses on real-world agent system builders.