Long awaited update: idea dashboard, my thoughts and where this is going in the near future

In my last post, I talked about imparting a trading idea process into the agent. Between then and now, I realized a prerequisite to giving an agent a proper trading idea process is actually a whole system for trading agents. This is because even though you can spend a lot of time perfecting one agent, that one agent probably won’t scale, so you actually need to make things very modular to scale which means setting up a system. When creating my system, here’s what I was thinking: 1 it should be auditable (you can see what the agents did, their reasoning, etc), 2 it needs to scale, 3 it has to actually work (obviously).

My first realization is that rather than running agents over a whole sector like I was previously, it’s actually better to run an individual agent over 1 ticker. So if I can perfect the process of researching and getting an idea on 1 ticker, then that will scale because I can simply just multiply the tickers I feed to agents and then I get coverage over what I care about. My next realization was that I didn’t even know how a bunch of agents getting trading ideas should even work. One huge problem is that two agents can actually propose conflicting ideas. For example, let’s say agent 1 makes an assumption about x when proposing an idea for ticker 1, then if agent 2 makes an assumption about x when proposing an idea for ticker 2, if those assumptions don’t match, then our system is going to output contradictory ideas. That realization served as the baseline for the whole system that I set up. Now, each agent runs over an individual ticker, but they share a database of facts and assumptions. Every idea links to assumptions made for that idea, and every assumption links to facts and their explanations for the links across the chain.

To get a bit more specific: assumptions are a topic and a view/belief on that topic. Each assumption that goes into the database should be orthogonal to the others in their topic. So for any individual topic, such as Fed Funds Rate October 2026, there can and should only be one view in our database on that. Any agent that wants to use Fed Funds Rate October 2026 needs to use that assumption that other agents are using for their ideas. If an agent does more research and finds facts that contradict the assumption, they can update that assumption. Because in the database, this is the only assumption on that topic and all agents using it have linked it to their ideas, then other agents will see that the assumption updated and why it updated. This will then inform them as they go through their trading idea process and the update will be reflected across the board as we get more information and validate or contradict our assumptions.

I also want to add a little color on the actual implementation because that’s just an idea, but doing this in practice also isn’t super easy… well the design isn’t. I had been planning with a coding agent and we were going to add a bunch of tags to every idea, assumption and fact so that we could find assumptions that are relevant to what an agent is trying to do even if they aren’t necessarily using that assumption. So like all fed assumptions would show up for any fed agent. All oil assumptions for any oil agent. As it turns out, that system is quite difficult and brittle. My original idea was to actually have a “librarian” agent to manage the database and that’s what I ended up going with. This has an issue where as the database gets larger, the librarian agent might have too much to look through. I realized that we can just create a system for a particular set of tickers, then multiply that overall system but keep them separate for other tickers. Thus, any individual librarian agent won’t actually see everything, they’ll just see what’s in their system. Then at some point I will create something to manage all of the smaller systems. Basically, it’s an AI pod shop hahaha. Every pod will be internally consistent, and I’ll deal with managing all of the pods when I get there.

After setting that up, I set up a VPS and am now running agents over a list of Kalshi tickers I specify at 6 hour intervals which will then update the website. As of literally the moment I’m writing this blog, there’s only 3 tickers in the list and only 1 getting published. I think that’s because the agents didn’t get actionable ideas on the other 2 / they didn’t find what they believed to be edge. I’m going to expand the list at some point in the near future depending on my interests. Right now, I think the building blocks are in place and I can go back to refining the trading idea process for individual agents now that the actual system is set up now.

As a side note, I’ve also done some more experimenting with other agents, an idea I think is very doable is a quant agent that researches systematic strategies. This is particularly appealing because the outputs are all entirely verifiable just by checking the sharpe out of sample and the real world explanation for why the strategy may work. Additionally, agents are just LLMs with loops, so this means that now they can try out a bunch of different ML models or read a bunch of papers and just keep trying things out until they find stuff that actually works. There is a bit of a problem in that if you try a bunch of stuff out, at some point you get false positives (p hacking kinda), but isn’t this what quants were doing anyways? Anyways, I know this idea should also work because there are some startups that are literally working on that and they’ve already been trading and are profitable.

Lastly, I’ve been so engrossed in coding in my free time I haven’t kept up as much with the markets. So now that my system is actually set up (mostly), I think I can finally return to writing about the markets too.