The Good Tech Companies - A Venture Studio Using Multi-Agent AI Systems to Make Better Investment Decisions

Episode Date: February 18, 2025

This story was originally published on HackerNoon at: https://hackernoon.com/a-venture-studio-using-multi-agent-ai-systems-to-make-better-investment-decisions. By far, the most compelling use case for Generative AI isn’t replacing human workers but rather vastly increasing the dimensions that people can explore. Check more stories related to machine-learning at: https://hackernoon.com/c/machine-learning. You can also check exclusive content about #artificial-intelligence, #super-vision-studio, #ai-driven-venture-studio, #tim-shea, #generative-ai, #multi-agent-ai-systems, #multi-agent-system-ecosystem, #good-company, and more. This story was written by: @missinvestigate. Learn more about this writer by checking @missinvestigate's about page, and for more stories, please visit hackernoon.com. By far, the most compelling use case for Generative AI isn’t replacing human workers or automating strictly defined workflows, but rather vastly increasing the dimensions that people can explore when making a decision.

This audio is presented by Hacker Noon, where anyone can learn anything about any technology. A venture studio using multi-agent AI systems to make better investment decisions, by Miss Investigate. This article was co-written with Tim Shea of Supervision Studio. By far, the most compelling use case for generative AI isn't replacing human workers or automating strictly defined workflows, but rather vastly increasing the dimensions that people can explore when making a decision. In this way, LLMs and diffusion models are incredible brainstorming partners. We can instruct LLMs about the knowledge domain we'd like it to represent and then precisely define how we'd like to see the output. For example, we could ask greater than, I want to create a database of the luxury retail holding companies like VF greater than corporation, New Guard Group, Only the Brave, and LVMH,

as well as the brands greater than they've recently acquired. Can you please generate an Excel spreadsheet with greater than the following columns, holding company, brand name, acquisition amount, greater than acquisition date, rationale for acquiring, queries like this are now considered table stakes, a staggering acquiring, queries like this are now considered table stakes, a staggering upgrade over traditional search engines that now seem quaint in the era of generative AI. However, some decision processes can be exceedingly complex. Sometimes we'd like to be able to model decision-making that might occur between multiple specialized personas, such as a CEO, CMO, or CTO, or specialized modules, such as an API connector, K1 parser, or SWOT analyzer. In the real world, we'd take the

results of that Excel spreadsheet and run it through a gauntlet of human experts. Strategists, investors, and domain experts would chime in, would construct what-ifs about the acquisitions, we'd dig into the P&Ls, and we'd dream up go-to market plans for potential future acquisitions. But thanks to modern LLMs, we can now create individualized AI agents that can model each of these behaviors and simulate having a room full of experts who inform our decisions. The Supervision Multi-Agent Operating System. So when we first founded Supervision Studio, an AI-accelerated venture studio focused on creating future brands, we immediately turned to multi-agent AI systems to model the proprietary decision-making processes that we call our

Supervision OS. With it, we can scour massive business databases in search of weak P&L models where AI can deliver outsized improvements. We can task the agents in our OS with hypotheticals about retail, CPG, print, entertainment, or commerce. The agents can crunch numbers, pontificate, and debate with each other. The agents can even produce deliverables that other agents criticize and iterate on. Multi-agent systems have a number of advantages over simple LLM chat interfaces. Each agent in the system can be trained with an enormous amount of specialized knowledge, the agents can be architected so that specific types of back and forth and checks and

balance occur. The agents can be strung together with deterministic modules, like a Python script, that perform specialized tasks. The agent system can be given an almost infinite compute space to iterate and come up with options. The agent system can present us with a dazzling array of options, each of which contains extremely well-crafted detail. The multi-agent system ecosystem. We also have a huge array of off-the-shelf options for standing up customized multi-agent systems like Autogen, Crew AI, and Langchain. We chose Autogen because it's free, there are tons of introductory videos, and they provide a sample code for multiple different multi-agent architectures that can be configured and stacked together. Autogen catalogs an extensive list of examples of what

it calls conversation patterns, such as two-agent chat, sequential chat, group chat, and nested chat. These patterns or architectures are an important part of the design process of the agent system. It's similar to how you'd build a team of human experts and predefine the rules of engagement between them. Different designs can produce wildly different outcomes. The simplest pattern that Autogen provides is a two-agent chat, where we define each agent's persona, where one agent simply asks the other agent the original prompt question, and where the agents go back and forth on a problem a set number of times. It's easy to set up, and we can read each interaction within their conversation verbatim. The sample code is below the output is pretty useful, and it gives us humans some ideas of what to do next. But it's not significantly more valuable than simply asking chat GPT the same

question in a single prompt. A slightly more advanced pattern is a sequential chat. In this case, we set UP multiple two-agent chats, where the output of each two-agent chat is passed onto multiple downstream two-agent chats. Think of this design like a CEO handing a task to a lower-level employee who makes decisions and then passes the decisions through multiple levels of approval. Sample code for this design is below. This pattern produces significantly more pointed and detailed feedback now that it's been vetted by a gauntlet of experts. And it emphasizes the importance of designing the right agent system for the right task. Autogen provides examples for many different design patterns, including group

chats and nested chats, but the field is incredibly nascent, T and the potential designs and architectures are limitless. We've looked at proposals and white papers for Magentic One, a multi-agent system from Microsoft where agents, like a web surfer or coder, work together under the guidance of a central orchestrator to solve complex tasks. Human personas. A Stanford-associated project where agents were trained to behave like the real-life subjects of qualitative interviews and then asked to converse and make decisions across a variety of topics. Project SID. A large-scale simulation of agent societies, where agents concurrently participated in jobs, trade, decision-making, and developed social systems, simulating human-like communities.

It's a fascinating field and we're actively experimenting with new frameworks and design patterns to provide rocket fuel to our decision-making and operational capabilities at Supervision Studios. If this is an interesting field for you, please reach out to us. We'd love to engage with other smart practitioners in the field. About Supervision Studios Supervision is an AI-driven venture studio focused on building future brands. For more information, please visit https colon slash slash supervision studio or contact Tim at supervision studio n n n. Thank you for listening to this Hacker Noon story, read by Artificial Intelligence. Visit HackerNoon.com to read, write, learn and publish.