For five years at One Mission Society, I ran cluster-based Kubernetes infrastructure.
The diagrams from that period are still on my computer. They show GKE and bare-metal nodes, load balancers, ingress rules, persistent volumes, databases, backups, monitoring, and separate production, staging, and sandbox environments. Each box had a job. Each connection carried a dependency. If something failed, the diagram helped us understand where to look for the issue and who needed to respond.
I spent much of my career working this way: making complex technical systems understandable enough to operate. Before Kubernetes, I worked in information architecture, search, web platforms, and enterprise applications. Afterward, I moved further into research systems, knowledge repositories, and AI-enabled workflows.
From the outside, my recent focus on qualitative research might look like a turn away from infrastructure.
It does not feel that way from the inside.
The object of care changed. The discipline did not.
In infrastructure work, I was concerned with how a workload moved through a system. Where did it enter? What depended on it? What state had to persist? What happened when a service failed? Could we restore the system without losing what mattered?
In qualitative research, I find myself asking a parallel set of questions about meaning. Where did an interpretation come from? Which words or moments in a conversation support it? What changed as it moved from transcript to concept, from concept to pattern, and from pattern to a decision? When AI participates, can a researcher still inspect that lineage? Can we recover when the analysis wanders?
The analogy has limits, and those limits matter. People are not workloads. A conversation is not a data packet. Human thought should not be made efficient by flattening it into something a system can route more easily.
But the system around interpretation can be designed.
That distinction has become central to my work.
Why thinking styles feel timely
Over the past year, I have been studying Indi Young’s approach to deep listening and thinking styles. A thinking style is not a personality type, demographic segment, or label assigned to a person. It describes a recurring way people reason, react, and orient themselves while pursuing a particular purpose.
Two people can make opposite choices for the same underlying reason. Two people can make the same choice while reasoning in completely different ways. Thinking styles help researchers look beneath visible behavior toward the interior logic shaping it.
That is not a new idea invented for the AI era. Indi has developed this work over decades, and I want to be precise about that lineage. She does not use AI agents to produce thinking styles. The agent experiments are my extension, not part of her method.
Still, thinking styles feel unusually suited to this moment.
Generative AI is very good at producing a clean summary of what appears to be common across a body of text. That ability is useful. It is also dangerous. The cleaner the synthesis sounds, the easier it becomes to miss the distinct ways people arrived at what they said.
Much of the current pressure around AI rewards convergence: find the pattern, compress the evidence, produce the answer. Thinking-styles research asks us to remain with divergence longer. It looks for meaningful differences in how people frame a purpose, what they pay attention to, which internal rules guide them, and what makes a decision feel possible or impossible.
I hesitate to call thinking styles the “trendiest” form of qualitative research. Trends happen at the surface. Thinking styles work below it. But I do think they are becoming more relevant as organizations realize that faster summaries do not automatically produce deeper understanding.
If AI makes behavioral patterns easier to detect, the value of research may move toward what behavior alone cannot explain.
Traceable and accountable
The phrase I am testing is agent-assisted qualitative analysis.
I prefer it to agentic analysis, at least for now. “Agentic” can suggest that an autonomous system owns the analysis. That is not the system I want to build. The surrounding infrastructure may be agentic; the analysis should remain agent-assisted.
I am interested in giving AI agents bounded analytical responsibilities. One agent might preserve the most empathic reading of a participant’s words. Another might challenge weak assumptions or identify where an interpretation reaches beyond the evidence. A third process might compare those perspectives. Deterministic rules can control which source material they read, where they write, what format they return, and when a human researcher must review the work.
The agents can create useful friction. They can keep alternate interpretations alive. They can help one researcher examine more material without pretending that volume and understanding are the same thing.
But they do not become the accountable researcher.
This is where my infrastructure history returns most clearly. In a reliable cluster, components do not all write wherever they want. Responsibilities are separated. State has an owner. Changes leave traces. Backups exist because failure is expected, not because the system is poorly designed. Monitoring makes the invisible visible before an incident becomes irreversible.
I am applying those lessons to the architecture surrounding qualitative analysis: separate roles, explicit handoffs, versioned source material, visible transformations, evaluation fixtures, and human review gates.
The purpose is not to make interpretation deterministic. It is to make the path around probabilistic interpretation traceable and accountable.
That is a subtle but important difference.
What cannot be automated away
There is a temptation to describe a group of agents as a substitute research team. I understand the appeal. Thinking-styles analysis is collaborative, time-intensive work. Independent researchers and small organizations may not have a full team available. I have felt that constraint myself.
But calling agents a team does not make them one.
Human researchers do more than generate multiple answers. They advocate for their understanding of a participant. They notice when a plausible grouping feels wrong. They negotiate language. They revise their own understanding while listening to someone else. Sometimes a discarded interpretation becomes as valuable as the accepted one because the discussion reveals why an apparent similarity did not hold.
An agent system can preserve several perspectives and make disagreement visible. It may even help a human researcher prepare for a stronger analytical conversation. It cannot quietly inherit the relational responsibility of listening to another person.
So the question I am pursuing is not, “Can agents automate thinking styles?”
It is, “Where can agents strengthen the analytical process without taking ownership of meaning away from the researcher?”
That question is less dramatic than full automation. It is also more difficult to answer.
It requires us to identify which parts of analysis benefit from computational assistance and which parts depend on human judgment, dialogue, and accountability. It asks us to design for uncertainty rather than treating uncertainty as a defect the model should remove.
The infrastructure remains
I did not leave infrastructure behind when I moved toward qualitative research.
I carried forward a concern with reliability, provenance, handoffs, and failure. I carried forward the habit of drawing the system so that hidden dependencies become visible. I carried forward the belief that complexity does not become manageable because one tool claims to contain it.
What changed was my understanding of what a system must protect.
In cloud infrastructure, we protected availability, data, and continuity. In research infrastructure, we also have to protect context, difference, and the relationship between evidence and claim.
A restored server can return to its last checkpoint. Human meaning is not so easily recovered once it has been compressed into the wrong category.
That is why I am drawn to thinking styles now. They resist the average user. They resist the convenient segment. They ask researchers to preserve different interior logics long enough for those differences to become useful.
And that is why I am experimenting with agent-assisted qualitative analysis. Not because I want AI to finish the research for me, but because I want to understand whether carefully bounded agents can help a human researcher hold more complexity without surrendering responsibility for it.
For years, I designed systems that kept machines coordinating under pressure.
Now I am trying to design systems that help people interpret without flattening one another.
The infrastructure is still there.
It is simply being asked to protect something more human.