Project · AI Automation

AI Stock Analyst: a multi-agent n8n workflow with a Telegram bot

A project built by Webziper. You message a Telegram bot, and an n8n workflow sends each stock through specialist AI agents for fundamentals, news and SEC filings, plus plain code for the technical numbers. A final decision agent applies hard rules with veto power and posts a BUY / WAIT / REJECT report back to Telegram.

Stack

n8n, Flowise LLM agents, Telegram Bot API, Finnhub, Yahoo Finance, SEC EDGAR

Type

Multi-agent workflow

Interface

Telegram

What does it do?

Most AI agent demos are one big prompt doing everything: reading the data, doing the math, making the call. When the answer is wrong, you can't tell which step failed. We built this project to test the opposite approach. Several narrow agents each look at one source, code handles anything calculable, and the final call sits under rules a human wrote.

Every ticker goes through six stages, one ticker at a time:

1. Fundamentals agent. Reads financial metrics from Finnhub and returns ACCEPTED or REJECTED, with its reasoning.
2. News sentiment agent. Reads recent news and rates the risk. It can issue a VETO.
3. Technical layer. Weekly and daily RSI, calculated in code.
4. Extra context. Earnings dates, insider trading, short interest, volume ratio and market cap. Also code.
5. SEC agent. Reads the company's XBRL filings from SEC EDGAR and summarises balance-sheet and debt risk.
6. CIO decision agent. Gets everything above and answers BUY, WAIT or REJECT in one sentence, under strict rules.

How does the Telegram bot work?

A Telegram bot is an account that hands every message it receives to your server. n8n's Telegram Trigger node receives those messages, so anything you send the bot starts the workflow. There's no dashboard to build or log into.

• /start runs screener mode. n8n pulls tickers from Yahoo Finance screeners, including Undervalued Growth Stocks and Undervalued Large Caps, merges the lists and drops anything below a market-cap threshold. Only tickers the fundamentals agent ACCEPTS get the full analysis.
• A single ticker like AAPL runs single-ticker mode. That company goes through the whole pipeline whatever the fundamentals verdict is. If you asked about it, you get an answer.

A Switch node routes each message, both paths merge into the same pipeline, and a Split In Batches node feeds the tickers through one at a time. Each agent only ever sees one company, and every ticker gets its own report.

Each report starts with the ticker and market cap, the verdict and a one-line reason, then lists the evidence behind it:

How do the agents split the work?

The rule is simple: each agent gets one source and one question. Only the last agent sees everything.

• Fundamentals agent (a Flowise flow that n8n calls). Finnhub's metrics are numbers, but interpreting them isn't arithmetic: a ratio that's alarming for one company is normal for another. The agent answers one question: does this business pass a fundamentals check?
• News agent. Recent news is formatted into a compact digest and the agent rates the risk. If the news points to serious risk, it issues a VETO that the final decision has to respect.
• SEC agent. The workflow looks up the company's CIK (its SEC identifier), downloads the XBRL filings from EDGAR and parses them. The agent only answers how risky the balance sheet is, especially the debt.
• Technicals and context are code. RSI is calculated, not guessed, and a parser turns it into a signal and a strength. An LLM would add nothing here except the chance of getting the arithmetic wrong.

How is the final decision made?

Ask an LLM "should I buy this?" and you'll always get a confident answer. Without constraints, that answer is a guess in good grammar. So the CIO agent works under rules it can't argue its way around:

• News shows high risk, or the news agent issued a VETO → REJECT, even if RSI is oversold.
• The SEC report shows high debt risk → REJECT.
• RSI above 70 → WAIT (overbought).
• RSI below 35, fundamentals ACCEPTED, and news and SEC risk low or medium → BUY.

The first rule matters most. A stock whose price has collapsed often looks cheap on RSI, and sometimes it's cheap for a reason. A news veto closes off that value trap with one line of logic. When no hard rule applies, the CIO agent defaults to WAIT: a BUY only happens when every BUY condition is met.

Why LLMs + code + hard rules?

The payoff is auditability. Every REJECT traces back to a specific rule and a specific agent's finding, so you can check the reasoning instead of trusting it.

What we had to solve

One filter, two entry modes

"Pass only tickers the fundamentals agent ACCEPTED" is right for screener mode, but it breaks single-ticker mode: ask about a company with weak fundamentals and you get nothing back. The fix is one filter with OR logic. It passes the item if fundamentals are ACCEPTED, or if it came from a single-ticker message. That only works if each item still remembers where it came from after the two paths merge.

Tickers with no news

Some tickers have no recent news; Realty Income (O) is a good example. In n8n, a node that returns zero items stops that branch by default, so a ticker can quietly drop out of the batch. The right handling is to treat "no news" as data: the news step always outputs an item, and the agent reads it as not enough information (neutral, no veto) rather than "low risk".

Where does this pattern fit your business?

Take the stocks out and the architecture is general: specialist agents judge one source each, code handles the numbers, and a decision agent applies rules your team agreed on, with vetoes for what must never slip through.

We build these for businesses as n8n automations and AI agents, with a human in the loop wherever a decision is sensitive. If your team makes the same kind of decision every day from several sources, tell us about it. We'll tell you honestly whether it needs agents, plain automation, or neither.