A production-quality, multi-agent buy-side financial research system powered by Claude Opus 4.6 with adaptive thinking and SEC EDGAR data.

Five coordinated AI agents mimic a real research function:

1. Agent 0 — Project Manager: CLI interface for project setup and orchestration
2. Agent 1 — Research Agent: Fetches SEC EDGAR filings (10-K, 10-Q) with cache-first logic
3. Agent 2 — Analyst Agent: Extracts compact CompanyFacts + CompanyBrief + QuoteBank per company
4. Agent 3 — Lead Analyst: Writes per-company analyst reports and a sector synthesis from compact summaries
5. Agent 4 — Visualization Agent: Generates professional charts (bar, line, comparison)

All outputs are cached. Arize Phoenix provides local observability at http://localhost:6006.

cd "Buy Side Financial Analyst"
pip install -r requirements.txt

export ANTHROPIC_API_KEY=your_key_here
# or copy .env.example to .env and fill it in

python main.py

$ python main.py

What would you like to do?
  [1] Start a new project
  [2] Continue an existing project

Enter 1 or 2: 1

── NEW PROJECT SETUP ──────────────────────────────────────
Project name: AI_Capex_2025
Companies: Microsoft Corporation:MSFT, Apple Inc:AAPL
Research question: Which company has stronger AI capex growth?

Financial Files/
└── MSFT/10K/, MSFT/10Q/, AAPL/10K/, AAPL/10Q/    ← raw filings (cached, not read by Agent 3)

Financial Data/
├── MSFT_facts_latest.json    ← structured KPI table + citations
├── MSFT_brief_latest.md      ← 800-1500 token company summary
├── MSFT_quote_bank.json      ← top 5-10 management quotes
├── AAPL_facts_latest.json
├── AAPL_brief_latest.md
└── AAPL_quote_bank.json

Financial Analyses/AI_Capex_2025/
├── analyst_reports/
│   ├── microsoft_corporation_analyst_report.md
│   └── apple_inc_analyst_report.md
├── sector_reports/
│   └── sector_report.md
└── visualizations/
    └── *.png

Agent 2 would read all raw SEC filings (up to 40k chars each) for all companies into a single LLM session. With 3+ companies × 2 filings each, sessions regularly hit the 200k context limit before the agent finished writing output.

Agent 3 then read those large markdown files for cross-company synthesis — compounding the token pressure.

Agent 2 — One Company at a Time

Each company runs in its own isolated LLM session (~35k tokens max vs. 200k limit). The session reads at most 2 filings (1 × 10-K + 1 × 10-Q) and produces three compact files:

Agent 3 — Compact Inputs Only

Agent 3 reads only the compact summary files — never raw filings. Its full context for N companies is predictable and small:

N × (facts ~3k + brief ~1.2k + quotes ~1.5k) + system + output
= 5 companies × 5.7k = ~28k tokens of inputs

When a specific citation is needed, search_excerpts(ticker, "keywords") retrieves only the matching paragraphs (~1–3k tokens), not the entire filing.

Arize Phoenix runs locally and auto-instruments all Claude API calls:

http://localhost:6006

View per-agent traces, token usage per company session, tool calls, and thinking budgets.

Set LLM_PROVIDER in .env to switch providers with no code changes:

See .env.example for a full template.

• Cache-first: Agent 1 never re-downloads filings < 24 months old
• No fabrication: All agents have explicit guardrails against hallucinated data
• No internet in analysis: Agents 2–4 read only from local files
• Isolated sessions: Agent 2 runs one LLM session per company to prevent context overflow
• Compact intermediates: Agent 3 reads only the pre-extracted summaries (facts + brief + quotes)
• Targeted retrieval: search_excerpts provides citation support without full filing ingestion
• Adaptive thinking: Agents 2 and 3 use extended thinking (Anthropic only) for deeper reasoning
• Configurable LLM: Switch between Anthropic, OpenAI, Gemini, or any local server via .env
• Observability: Every API call is traced via Arize Phoenix

See ARCHITECTURE.md for the full system design and data flow diagram.