What is Agentic AI in Finance? Meet Mycroft

A chatbot that answers financial questions is useful. An AI system that gathers the data, reasons through it, takes action, and improves from what happened is something else entirely. That distinction is the starting point for Mycroft, a project exploring what agentic AI actually looks like when applied to investment decisions rather than just financial Q&A.

What makes AI "agentic"

Agentic AI is described here as AI that takes the initiative to solve complex problems through sophisticated reasoning, rather than simply responding to a prompt. Unlike a regular chatbot, an agentic system perceives what's happening, reasons through it, takes action, and keeps learning, handling financial tasks with minimal human intervention at each step. The difference isn't cosmetic: a chatbot gives advice, while an agentic system is built to actually do something with that reasoning.

The four-step loop: perceive, reason, act, learn

Agentic AI in this project works through four connected steps. It perceives by gathering data from market feeds, financial databases, and regulatory filings, functioning like a team of research assistants working around the clock. It reasons by using large language models as the coordinating layer, understanding tasks and directing specialized models toward different financial functions. It acts by connecting to financial tools to execute tasks based on its plans, moving past simply giving advice. And it learns by improving through feedback loops that make the system smarter over time. Each step depends on the one before it, which is what separates this from a single-shot chatbot response.

Why the name Mycroft

The project takes its name from Mycroft Holmes, Sherlock Holmes's older and, in the stories, even more brilliant brother. It's an open-source educational experiment in AI-powered investment intelligence, built around the idea of "using AI to invest in AI," and explores how specialized agents can analyze the AI sector while sticking to disciplined investment strategies rather than chasing every signal that appears.

A small band, not a full orchestra

Rather than building out hundreds of financial agents from the start, the project deliberately begins with just a handful of core agents plus the Mycroft coordination layer. That restraint is the point: starting small lets the team see what actually works in simulated trading environments instead of relying on theoretical assumptions about what should work.

The initial roster includes analytical agents that act as financial detectives gathering information about AI companies, portfolio agents that turn that knowledge into actionable investment strategies through simulated trades, advisory agents that handle conversational financial advising for human interaction, and intelligence agents that monitor news, social sentiment, and regulatory developments in real time. Sitting above all of them is the Mycroft layer itself, which functions like a conductor: it handles cross-agent validation to resolve conflicting conclusions between agents, dynamic task allocation to redistribute resources when market priorities shift, and pattern recognition to find connections across developments that might otherwise look unrelated.

For now, the project is deliberately limited to four components: one analytical agent focused on earnings reports, one portfolio agent testing an investment strategy, one intelligence agent monitoring industry news, and the Mycroft layer coordinating them. New agents only get added once they prove their worth in the simulated environment, rather than being built out speculatively.

Key takeaways

• Agentic AI is distinguished from chatbots by a four-step loop: it perceives data, reasons through it with LLMs, acts using financial tools, and learns from feedback over time.
• The Mycroft project starts with a small core of specialized agents (analytical, portfolio, intelligence, advisory) rather than building a large agent roster upfront.
• A dedicated Mycroft coordination layer handles cross-agent validation, dynamic task allocation, and pattern recognition across the other agents.
• The project is explicitly educational and open-source, testing strategies in simulated trading environments before scaling based on what proves out.

Who this is for

This is aimed at anyone curious about how agentic AI concepts apply to a real domain like finance, especially students and builders interested in open-source, education-first AI projects. The project is led in collaboration with Professor Nik Bear Brown, and anyone interested in learning more or joining is pointed to the Humanitarians AI website and YouTube channel.