AI for Investing
When you need an agent, not a prompt
A question I get constantly: should I use an agent for this? Usually, no. Not because agents are not capable, but because most tasks do not need that capability, and using an agent when a simpler thing works is driving a Formula 1 car to the grocery store. Think of the options as a ladder with three rungs.
The three rungs
Level 1, a single tool call. One question, one tool, one answer. "What is Apple's current price?" The user asks, the tool fires, the answer comes back. Most finance questions that seem simple actually are. "What is the P/E?" "What sector is this?" These do not need an agent. They need a tool call.
Level 2, a workflow. Several steps in a sequence you defined in advance. "Compare JPMorgan and Goldman on valuation" means fetch the fundamentals for one, fetch them for the other, compare, format. The model is not doing much: it might parse the request at the start and write the output at the end, but the sequence is fixed. When you know the steps, you do not need an agent to work them out. Workflows are underrated. Most "I need an agent" requests are really "I need a workflow."
Level 3, an agent. The agent decides what to do next based on what it finds. "Analyse my portfolio and identify the biggest risks." It might fetch fundamentals for all three holdings, notice they are all tech, decide to check correlations, find two of them move together, decide concentration risk is worth computing, and pull it into a report. You did not specify those steps. The agent chose them.
Can you write down the steps before running? If yes, you probably don't need an agent.
If yes and it is one step, use a single tool call. If yes and it is several, a workflow. Only if no, where the task has to adapt to what it finds, an agent. Even then, write down the steps you expect it to take. An agent can deviate from the plan, but having one helps you choose the tools and notice when it goes somewhere odd.
Each rung up costs more
A single tool call is one model call and one tool call. An agent analysing a portfolio might make ten to fifteen model calls as it works through the problem, and it can take a different path the second time. None of this is a reason to avoid agents. It is a reason to use them only when you need them. The impressive agent demos often solve a problem a well-designed workflow handles better.
The complexity trap
The trap I see constantly: "I want to build an AI portfolio assistant", and the person immediately starts building an agent. Break "portfolio assistant" down. "Show me my positions" is level 1. "What is my portfolio worth today?" is level 2. "How has it performed this month?" is level 2. "Generate my monthly report" is level 2. "Should I rebalance?" is level 3. Four of five features are level 1 or 2. Build the workflows first. Add an agent only for the tasks that actually need one.
Keep going
Once you do need an agent, see what one is made of and the four patterns. Start at AI for investing for the whole picture, or get the full book, AI Agents for Investing. I am publishing more in the coming weeks. Subscribe to get it.