One agent, seven windows
What a stock is worth looks different through every window you read it from. Fundamentals show the health of the business. Technicals show where price and momentum sit right now. Macroeconomics shows the weather all of it happens in: yields, the dollar, volatility. An honest market-analysis agent reads all three at once and says how it weighed them when they disagree.
The financial-interactive agent in melchizedek is built exactly that way: one model working through seven narrow instruments. Each tool does one narrow job. get_company_fundamentals returns valuation and growth metrics for a single ticker; momentum indicators and moving averages come from get_technical_indicators; get_macro_metrics reads treasury yields, the strength of the dollar, and market volatility. Search fills the gaps left by structured feeds, capturing earnings surprises, filings, and sudden catalysts. Every function returns precise numbers or an explicit empty result. They provide raw data, leaving the opinion entirely to the synthesis.
synthesis is the intelligence
The work is in how the agent composes those tools. Before it touches any data it classifies the question: a volatility check is one function call, a full ownership analysis is nearly all of them. Classifying first is what separates an analyst working a method from a system that calls everything it has and sorts the pile out afterwards.
Once the data arrives, the agent applies an opinionated weight to each window. In this synthesis hierarchy, fundamentals and valuation are the anchor. They carry most of the decision, because they answer the question ownership actually turns on. Technicals refine that answer by setting timing and risk boundaries: reasonable entry points, and stops scaled to the stock’s own historical volatility rather than to a round number. Macro usually calibrates position size, and it holds one veto: when volatility spikes and liquidity thins, it can bar new entries whatever the other windows say.
The output contract asks for one kind of honesty: every verdict has to be falsifiable. The agent gives a precise call and, beside it, the price level or event that would reverse the call. Even a neutral stance names its condition, the earnings print or the drop that would force a reassessment. An agent that can say what would prove it wrong is one you can hold to an answer.
Here is the whole discipline in one turn: plan, look, weigh, commit.
This principle extends far beyond finance. Provide an agent with many narrow, truthful windows rather than one vague abstraction. Require it to plan its approach before it seeks information. Rank your sources before they conflict, because they will. Ask every answer to carry the conditions of its own reversal. Data becomes judgment only when something is at stake, and an agent worth trusting names the price or the event that would make it wrong.