How I work
How I work with AI agents
Each thesis on this site answers a research question I have worked through with AI agents. The agents do much of the legwork; I set the questions, review what they produce, and write the conclusions. Agents and I do this work in my proprietary research management system (RMS). What you read here is the published end of my research process: the thesis and the models behind it.
Start with a question that matters
Research starts when a topic becomes important to my view of the US economy. AI capital spending, for example, is now large enough that its path shapes the growth outlook, so it gets its own line of research. I work with agents to turn the topic into a well-framed research question whose answer would change how I see the broader economy.
Agents do the legwork; I direct and review
From there the work is iterative. Agents plan work, carry it out, and record their findings for me to review. I read what they produce, accept or reject it, and decide where they go next. The RMS is built so that agents can do this work well.
Nothing lives in a chat window that disappears when the conversation ends.
Any agent, the same memory, continuous research
Agents connect to the RMS through MCP, an open standard, so I can work with Claude, Codex, Gemini, Grok or any open-source model that supports it. Whichever agent I use starts from the same place. The RMS keeps a succinct record of my research: the question I'm trying to answer and my hypothesis. When an agent needs more information, it can search everything I have done: notes, reviewed agent research, and financial models. The RMS optimizes search via a ranking algorithm, so agents can find relevant information faster.
Evidence standards and data on hand
Agents are only as good as their sources. Mine follow written evidence standards. They go first to primary sources: company filings, press releases and investor-relations sites. The data they need is built into the RMS: SEC filings, earnings-call transcripts and market data can all be accessed via standardized calls to the RMS, with the data cached so that repeated questions don't overload the providers.
Bottom-up models behind the macro view
Agents build financial models for companies based on industry templates I originally created for their reference. Every model possesses automated checks that must pass before I review. The models are written in Python, which agents can build, test and update far more reliably than a spreadsheet. They are then rendered as interactive pages for me to read and review.
Modeling this way has more benefits than just being agent-native. One particularly useful benefit is that I can propagate macro scenarios through bull/base/bear cases in dozens of models, consistently and efficiently.
My time is spent on decision making
Because agents carry the desktop legwork, most of my time goes to judgment: deciding what the evidence means and what I believe.
Want the same setup for your team? Consulting →