One agent, seven windows

2026-07-11 · #tools #agents

No single data source knows what a stock is worth. Fundamentals know the business but not the moment. Technicals know the moment but not the business. Macro knows the weather but not the ship. An agent that answers market questions honestly has to look through many windows at once — and know which one to trust when they disagree.

The financial-interactive agent in melchizedek is built as exactly that: one mind, seven instruments. Each tool is a deliberately narrow function. get_company_fundamentals returns valuation and growth for exactly one ticker. get_technical_indicators returns RSI, MACD, moving averages, ATR. get_macro_metrics reads treasury yields, the dollar, the VIX. get_analyst_recommendations returns Wall Street’s consensus and price targets. Search fills what no structured feed provides — earnings surprises, filings, catalysts. Every function returns numbers or an explicit empty result. None returns an opinion.

searchhistorymacrofundsanalyststechsnewsfundamentals & valuation — the anchortechnicals — timing, sizing, stopsmacro — calibrates; can vetoa falsifiable verdict
Seven narrow, truthful windows; one opinionated synthesis. The weighting is the intelligence.

synthesis is the intelligence

The agent’s skill is not calling tools — it’s composing them. Before touching anything, it classifies the question: a VIX check needs one call; “should I own this?” needs nearly all of them. Discipline at this step is what separates an analyst from a slot machine that pulls every lever every time.

Then the windows get weighted, and the weighting is opinionated. In this agent’s synthesis hierarchy, fundamentals and valuation are the anchor — 80 to 90 percent of the decision, because they answer should I own this business at this price. Technicals never overturn that answer; they refine its timing and set its risk — where the entry is reasonable, where the stop belongs, sized by actual volatility rather than a round number pulled from air. Macro normally just calibrates position size, but it holds one veto: in a genuine storm, it can bar new entries entirely.

The output contract enforces honesty. Every verdict must be falsifiable — not “looks interesting,” but a named call with the exact level or event that would change it. Even “no edge” must state its trigger: becomes a buy below this price, reassess after this earnings print. An agent that cannot tell you what would prove it wrong is not analyzing. It’s performing.

Here is the whole doctrine in one turn — plan, look, weigh, commit:

┌─ trace: a full workup (ticker fictional, figures illustrative) ─ interactive

That is the general lesson, and it travels far beyond finance. Give an agent many narrow, truthful windows rather than one wide, vague one. Make it plan before it looks. Rank its sources before they conflict, because they will conflict. And require every answer to carry the conditions of its own reversal. Data becomes judgment only when something is at stake in being wrong — and a well-built agent is wrong in public, precisely, where the next query can hold it to account.