How I built a 12-agent stock portfolio research desk at $0.03 per report
If you hold a dozen individual stocks and ask a single-prompt LLM chatbot how a company's quarter went, two things usually go wrong. First, the model has no context on your actual portfolio weights, sector exposure, or s
If you hold a dozen individual stocks and ask a single-prompt LLM chatbot how a company's quarter went, two things usually go wrong. First, the model has no context on your actual portfolio weights, sector exposure, or single-position risk limit. Second, a single prompt tends to mirror the framing of your question and summarize headlines without checking primary SEC filings.
Open-source multi-agent frameworks like TradingAgents showed a better structure: split equity research into specialist roles and force an explicit debate before anyone writes a conclusion. The catch is that cloning a Python repo, wiring market data keys, and running terminal scripts for one ticker at a time is too much friction when you want daily coverage across an entire portfolio.
So I built Volamind, a web app that runs a 12-agent research desk on the stocks you own, checks concentration against your own rules, and grounds follow-up chat in SEC EDGAR filings and live quotes.
You can explore a full Microsoft report, the bull-vs-bear debate, and cited follow-up chat with zero login here:
Inspect the live Microsoft report on /demo
Splitting one prompt into 12 structured agent steps
Rather than asking one model to do everything in one pass, each stock report runs as a directed graph of specialized steps:
- Four specialist analysts execute in parallel: a fundamentals analyst reading financial statements and SEC EDGAR filings, a technicals analyst reading price action and momentum, a news analyst scanning recent company and macro events, and a sentiment analyst.
- A bull researcher and a bear researcher take those four briefs and argue both sides of the thesis over multiple rounds. Having a dedicated bear agent prevents the sunny consensus tone you normally get from chat models.
- A research manager synthesizes the debate, a trader agent translates it into an actionable stance, and three risk analysts (seeking, neutral, and conservative) challenge the downside before the desk issues one dated card (
Overweight,Equal-weight, orUnderweight) with explicit key risks.
For follow-up questions, the companion chat calls structured tools (get_quote, get_market_backdrop, get_recent_filings against SEC EDGAR, and grounded web search) so factual numbers link directly to the underlying filing or quote rather than relying on model memory.
Solving the unit-economics trap before opening a no-card trial
Running 12 LLM steps per stock sounds expensive if you run it naively per user request. As a solo founder, my biggest constraint was making sure a traffic spike could never turn into a surprise cloud bill, while also avoiding token credit meters that make users hesitate before clicking anything.
Three architectural choices made the math work:
1. Shared ticker-day report cache
An equity research report on MSFT or NVDA for 2026-11-20 depends on the instrument and the market data as of that UTC date, not on who asked for it. Reports are keyed by (instrument_id, as_of_date).
When our nightly coverage job runs MSFT once, it costs roughly $0.03 in tokens. Whether 1 investor or 500 investors hold MSFT in their portfolios that day, everyone reads that dated report from PostgreSQL at zero marginal LLM cost. Portfolio-specific checks (like flagging that MSFT is 28% of your portfolio when your personal concentration limit is 20%) are computed deterministically on top of your holdings.
2. Global daily uncached-ticker budget
What if someone imports 40 obscure micro-cap tickers that nobody else on the platform holds? We enforce a global daily budget cap (60 brand-new on-demand ticker reports per day and 40 nightly scheduled reports prioritized by how many portfolios hold each ticker). Cached tickers remain instant and unlimited, while worst-case daily report token spend stays capped around $3/day.
3. Feature tiers instead of credit anxiety
Because report generation is shared and bounded, we keep pricing simple instead of charging per prompt:
- Free (
$0forever) includes the portfolio tracker, watchlists, sector and currency exposure breakdown, concentration check against your own position limit, side-by-side stock comparison (1-year return, annualized volatility, max drawdown), and an earnings, dividend, and macro calendar filtered to your holdings. - Every new sign-up gets a 7-day full Pro trial with no credit card required, unlocking full bull-vs-bear reports, daily Signals across all holdings, and cited SEC filing chat.
Honest trade-offs I am still refining
Two real edge cases showed up quickly in production:
- Daily rating sensitivity on borderline names: when a stock sits right on the boundary between
Equal-weightandOverweight, a noisy news day can flip the badge more frequently than a human analyst would. Adding thesis hysteresis (requiring conviction to cross a wider band before changing a prior rating) is next on my roadmap. - Financial-sector ratios: standard free-cash-flow margin comparisons work cleanly for software and industrials, whereas banks and insurers need book-value and net-interest-margin templates.
Try it on your own holdings
Volamind is for research and education only, not financial advice, and it never asks for brokerage credentials (you enter tickers and cost basis directly).
- No-login interactive demo: https://volamind.com/demo
- Create a free account (includes 7-day Pro trial, no card): https://volamind.com/sign-up
If you hold individual stocks or build multi-agent workflows, I would love your feedback on the bull-vs-bear debate structure and what data sources or screening checks you would want next!
Originally published by Dev.to WebDev. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.