Multi-agent research system Early 2026
Stock Advisor
A stock research assistant built on Google's Agent Development Kit. A portfolio-manager agent hands work to four specialist agents, then combines what they find into one weighted recommendation, streamed live into a Streamlit app. It's for learning and research, not financial advice.
The agents
A PortfolioManagerAgent sits on top and delegates to four sub-agents through ADK: a data fetcher (prices, volume, the 52-week range, and market context from the VIX and the major indices), a technical analyst, a fundamental analyst and a news-sentiment agent. Every data operation is a typed ADK FunctionTool, ten in all.
Five stages per stock
Each analysis runs in order. The technical analyst computes RSI, MACD, Bollinger Bands, moving averages, the stochastic oscillator, ATR and OBV, looks for patterns like golden crosses and support and resistance, and scores the stock from -100 to 100.
The fundamental analyst does the same with valuation, margins, growth, debt and analyst targets. The sentiment agent scores recent headlines with TextBlob and picks out likely catalysts.
The recommendation
The portfolio manager weights the scores: fundamentals 40%, technicals 30%, sentiment 15% and the wider market 15%. It turns that into a call from strong sell to strong buy, with an entry zone, a stop-loss at twice the ATR, price targets and ATR-based position sizing.
Memory and streaming
A session memory caches price data for five minutes and fundamentals for ten, lets the agents read each other's output, and keeps a history of past analyses and the portfolio. Results stream into the Streamlit app as they're computed, across six tabs: the agent run, technicals, fundamentals, news, the portfolio and an architecture diagram.
Without a Gemini key it falls back to a direct analysis engine that uses the same tools with rule-based synthesis, so it still runs.