What a stock analysis agent should do before it makes a claim
A stock analysis agent is a language model wired to tools: data lookups, screens, comparisons. The wiring is the interesting part. An agent that retrieves evidence first and writes second can support research. An agent that writes first and decorates the text with numbers afterward is a liability with good grammar.
If you are evaluating one, judge the behavior before the claim, not the confidence of the claim itself.
Five behaviors that separate research from fluency
**It names the source and the date.** Every figure in an agent's answer should be traceable: which filing, which price series, recorded when. An agent that cannot show its working has not done any.
**It refuses what it does not cover.** Ask about an obscure company and a trustworthy agent says "not covered" rather than assembling a plausible description from training data. The refusal is a feature. It tells you the covered answers are grounded.
**It keeps selection deterministic.** When an agent returns a list, the companies on it should be there because stored data met a stated rule, never because the model felt they fit. The model may explain the list. It should not be the thing that picked it.
**It states the rule.** "Oversold" should unpack to something exact, such as a 14-day RSI below 30. "Long dividend growth streak" should mean a counted number of consecutive years. Vague labels hide weak logic.
**It separates fact from interpretation.** "Revenue grew 40% in the year ending March 2026, per the annual filing" is a fact. "This company is well positioned" is an opinion. A careful agent marks the difference instead of blending the two.
How stocks-llm applies this
stocks-llm is constrained to a catalog of covered public companies. It does not invent a ticker, a price, or a figure; an uncovered company is reported as not covered. Figures carry sources and as-of dates, missing values are shown as `n/a` rather than estimated, and prices are delayed daily closes. Screens recompute nightly from stored data, and a company with an unevaluable series is excluded rather than guessed at. The about page documents the full data model.
The same discipline is exposed to other software. A public, read-only MCP endpoint offers three tools for searching companies, fetching a dated profile, and comparing companies, so an agent you build can pull the same verifiable data instead of paraphrasing a web page.
A concrete test you can run
Give any agent this question: which companies have the longest dividend growth streaks?
Then apply the checklist. Did it state how the streak is counted? Does each company carry dated evidence? Can you open a primary source from the answer?
For a baseline, compare the agent's answer against a deterministic version of the same question: the longest dividend-growth streaks screen counts consecutive non-decreasing per-share annual dividends derived from SEC XBRL filing data, and states that rule on the page. Where the agent's list and the deterministic list disagree, the disagreement itself is the finding. The research task that remains is yours either way: read the filings behind a streak and decide whether the payout is durable.
The limit that does not move
However good the wiring, an agent cannot verify the future. It cannot turn a screen into a forecast or a comparison into advice. Its honest ceiling is a well-sourced, dated picture of what is knowable now, plus an explicit account of what is missing. Agents that claim more are not analyzing. They are marketing.
stocks-llm is for informational research only, not financial advice. Verify material information independently against primary sources before making an investment decision.