• Hey everyone! Hope everyone’s doing great—doing more daily writing around AI and financial services on X nowadays—feel free to follow me there. Every week I’ll try to pick something that I’ll share with you all here via email.

• My full time job, Arcana Advisors, has been busy too. We spent the summer working with Tier 1 and 2 banks around helping build out complex AI-powered workflows and use cases. We’re working with some banks around helping with model fine-tuning and training. If you want to learn more arcana-advisors.com.

• Looking to test out a few projects in the AI and financial crime space at a pretty deeply discounted rates with fintechs or banks that work directly with customers. Hit me up!

Agent Chains & Prompt Chains

Whether its prompting or through agents, I find chains one of the most useful ways to structure production level AI Agents.

I was reminded of this when I was reading Anthropic’s guide to prompt engineering yesterday. It talks about some of the advantages, like increasing the visibility and accuracy of the output (guessing and approximating versus running calculations.) I never realized but Anthropic's guide suggests breaking prompts up into subtasks where Claude can give each tasks 100% of its attention. This makes sense, but in an agent framework these tasks could easily be delegated to subagents too.

Prompt chains are incredibly useful, but so are agent chains, which don't get enough press. Agent chains are...chains...of specialized AI agents. They all are a small part of a larger workflow and pass their output and context onward for more accurate results and higher visibility and explainability.

I first learned about agent chains when looking into AutoGen, one of the most popular agent frameworks. I wasn't too happy with most of the agent frameworks and felt like they did a lot more than I needed. Now I just use Claude Agent SDK.

My use case was around investment research. One of my first consulting projects was to build a daily scanner for family office portfolio manager. He wanted agents to do daily scans around my portfolio, and daily scans across the market to search for alpha. It didn't even really had to be good alpha, just something worthwhile—the end deliverable ended up being a personalized morning research note. I was more interested in taking it a step further: a chain of agents would run quantitative and technical analysis on the stocks I was more interested and give me a summary report and a buy/sell/hold signal.

It actually worked really well! Each of the agents were specialized in their respective skill i.e. the technical analyst was an expert in technical indicators. They would do their work and execute their output. It would give me an opportunity to review their work and input it into the next step of the process (this can definitely be automated, but I liked the oversight.)

Not everything will work, and a lot of things will change. For example, you probably don't need robust agents for each task; you can accomplish a LOT more with prompt chains now. My advice for when to use agent chains is when you need really high explainability—think compliance workflows. Another thing that didn't work was introducing hierarchy—it dramatically increased the latency and time it took to get to a output, and the output quality was marginal (though, you can see a world where agent chain workflows chain into each other. I just don't think getting more granular will yield better output or results.)

Now, most of my work around agents is just building off of this agent chain concept. It's scaled pretty well and can handle production level analysis and workflows. The latency can sometimes be an issue but frankly in most AI financial services use cases, spending 3 minutes waiting for an AI to do what takes a human 3-5 hours is a fine tradeoff.