The brief: earnings memo from a ticker

You type one command. The agent outputs a professional research memo.

$ python memo_agent.py INFY

======== EARNINGS RESEARCH MEMO ========
Ticker: INFY (Infosys)
Date: 2026-04-16
Analyst: AI Research Agent

PRICE ACTION:
- Current: ₹1,542.50
- YTD Change: +18.3%

KEY HIGHLIGHTS (Q1 FY27):
• Net revenue ₹22,110 cr (+2.1% QoQ)
• Operating margin 21.3% (stable)
• Digital services revenue +12.5% YoY

BULL CASE:
- Strong AI transformation pipeline
- Digital revenue acceleration expected
- Stable margins with operating leverage

BEAR CASE:
- Macro uncertainty in US IT spending
- FX headwinds on rupee strength
- Competition from high-growth startups

RISKS:
- Client concentration risk (top-10 = 45%)
- Rupee appreciation pressure on margins
- Tech talent acquisition costs rising

OPPORTUNITIES:
- GenAI consulting services (new segment)
- Cloud migration acceleration post-COVID
- India domestic market growth (10%+ CAGR)

RECOMMENDATION: BUY
Target Price: ₹1,750 (13% upside)
Thesis: Fundamental strength in digital transformation
        + moderate valuation = compelling entry.

Under the hood: The system fetches real stock prices (yfinance), latest earnings headlines (web search), PDF quarterly filings (RAG), then routes data through five specialized agents (Planner → Researcher → Writer → Critic → Publisher) with feedback loops. No hallucination. Everything cited.

Architecture: 5-agent orchestration with feedback

This is a full production system. Each agent has a role, a system prompt, and access to tools.

1. Planner
(Define research plan)
→
2. Researcher
(Gather data)
→
3. Writer
(Draft memo)
→
4. Critic
(Review & iterate)
→
5. Publisher
(Format & output)
📊
Real scenario
SouthBank India analyst desk (Mumbai)

SouthBank India's equity desk analyst needs 10 research memos per day for Indian bluechips (INFY, TCS, RELIANCE, HDFCBANK). Manually crafting each takes 3 hours. The agent produces one in 2 minutes—with better citations. The analyst reviews and publishes.

Tool definitions: four data sources

Each tool is pure Python—no external APIs beyond yfinance (free).

Tool 1: fetch_price(ticker)

Returns current price, change %, and 52-week range from yfinance.

Tool 2: fetch_filing_pdf(ticker, period)

Downloads the latest quarterly filing PDF (if available publicly), stores it locally.

Tool 3: search_recent_news(ticker, num_results=5)

Queries Google News (or DuckDuckGo, which requires no API key) for headlines, returns title + snippet.

Tool 4: query_filing(question, pdf_path)

RAG over the PDF: embed the filing, search for relevant chunks, pass them + question to LLM. Returns cited answer.

RAG vs. hallucination:
Without RAG, the LLM might invent earnings numbers. With RAG, every number is grounded in the actual PDF. The agent cites page numbers.

System prompts for each agent role

Every agent sees the same ticker data but has a different perspective. Each prompt is 1–2 paragraphs.

Planner prompt

"You are a research strategist. Given a ticker, outline a 5-step plan for building a research memo: (1) Confirm ticker and sector, (2) Identify key metrics from latest filing, (3) Surface recent news themes, (4) Extract bull/bear arguments from SEC filings + analyst consensus, (5) Draft a recommendation. Be concise."

Researcher prompt

"You are a financial researcher. Execute the plan: fetch price data, search news, skim the filing for revenue/margin/guidance, and list top 3 bull and 3 bear arguments with citations. Format: {price, change_pct, headlines: [title, snippet], key_metrics: {}, bull_args: [{arg, source}], bear_args: [{arg, source}]}"

Writer prompt

"You are a sell-side analyst. Draft a Morgan Stanley–style memo using the researcher's data. Include PRICE ACTION (2 lines), KEY HIGHLIGHTS (3 bullet points), BULL CASE (3 bullets), BEAR CASE (3 bullets), RISKS (3), OPPORTUNITIES (3), and RECOMMENDATION (BUY/HOLD/SELL + target price). Be authoritative. Every claim must cite a source."

Critic prompt

"You are a senior analyst. Review the draft memo. Check: (1) Are all claims cited? (2) Is the recommendation supported by bull/bear balance? (3) Are risks grounded (not generic)? (4) Is the target price justified? (5) Are there missing segments (e.g., ESG, capex)? Return JSON: {passed: bool, feedback: string, revisions: [{section, fix}]}"

Publisher prompt

"You are a formatting expert. Polish the memo: ensure headers are consistent, indent bullet points, format currency with commas, and return valid JSON matching the ResearchMemo schema. Do not alter content; only structure."

Pydantic output schema: ResearchMemo

Every agent outputs JSON. We enforce structure with Pydantic so no agent can omit a field or mislabel a section.

memo_schema.py Structured research memo output

Orchestration: the main loop with critic feedback

The choreography matters. Planner runs first, Researcher uses the plan, Writer uses research data, Critic reviews, Publisher formats. If Critic finds issues, Writer gets called again.

main_orchestrator.py 5-agent orchestration loop with feedback

Tool implementations: fetch_price, search_news, query_filing

Three practical examples. All are synchronous Python; no async complexity yet.

tools.py Data fetching tools: yfinance, DuckDuckGo news, PDF RAG

Streamlit deployment: web UI for the agent

Wrap the pipeline in a simple web app. Paste your ticker, hit submit, and stream live agent updates.

app.py Streamlit front-end for the memo agent

Evaluation: test on 3 known tickers

After training, evaluate on INFY, TCS, and RELIANCE.NS. Check these rubrics for each memo:

Rubric (5 points each):

1. Price data is current (matches yfinance) ✓
2. Highlights cite specific metrics (revenue, margin, growth %) ✓
3. Bull/bear cases are balanced (not 3 vs. 1) ✓
4. All claims are cited (no unsourced assertions) ✓
5. Recommendation (BUY/HOLD/SELL) is justified by bull/bear score

Portfolio pitch: how to present this on your resume

Project title: "Earnings Research Agent – Multi-Agent AI System"

Impact statement: "Built a 5-agent orchestration system (Planner/Researcher/Writer/Critic/Publisher) that produces Morgan Stanley–style research memos from ticker input. Integrated yfinance + PDF-based RAG to ground claims in actual filings. Deployed on Streamlit. Reduces manual analyst time from 3 hours to 2 minutes per memo."

Technologies: Python, Ollama (or Claude API), Pydantic, yfinance, sentence-transformers, PyMuPDF, DuckDuckGo News, Streamlit

Key learnings: Multi-agent orchestration, system prompts, RAG pipelines, feedback loops, structured outputs, production deployment

How real research firms use agents

Every major investment bank now uses multi-agent systems for research. Bloomberg, Goldman Sachs, and Morgan Stanley have internal versions of this memo agent. The architecture you've built—Planner → Researcher → Writer → Critic → Publisher—mirrors their actual workflows. The difference is scale (10k tickers, real-time updates) and integration (live Bloomberg terminals, SEC filing feeds). But the core pattern is unchanged: break the problem into roles, give each role a system prompt, route data through the chain, and use critics to enforce quality. You've learned the production pattern.

Multi-agent orchestration RAG Critic feedback loops Structured outputs Production deployment
🎯 Challenge: Add an ESG Analyst Agent

Task: Extend the pipeline to include a 6th agent: the ESG Analyst. After the Writer completes the memo, this agent appends an "ESG Assessment" section: (1) Does the company have ESG controversies? (2) What is its carbon intensity vs. peers? (3) Board diversity score? The ESG Analyst searches news for "INFY ESG", "INFY environmental scandal", etc., and returns a structured section that the Publisher incorporates into the final memo.

Hint: Create an `esg_analyst.py` agent with its own system prompt. Insert it after the Critic in the orchestration loop, before the Publisher. Have the ESG Agent search for ESG-related news and controversies, rate ESG risk (high/medium/low), and return a structured ESGAssessment Pydantic model. Update ResearchMemo to include an optional `esg_assessment` field. The Publisher then formats it into the final JSON.

Test your knowledge

Question 1 of 5
Why does the Critic agent need to run before the Publisher?
Question 2 of 5
What does RAG (Retrieval-Augmented Generation) solve in the filing research tool?
Question 3 of 5
What is the Planner agent's role?
Question 4 of 5
How does the Critic iterate if it finds issues?
Question 5 of 5
Why use Pydantic for the ResearchMemo output?
0
/5
Excellent! You've mastered multi-agent orchestration.
Africa & India's Best AI Academy
🤖 AI-first 🎓 Placement Assistance

You've built a research agent.
Now ship agents at scale.

Join the Best AI Program: The AI-Era Agentic Engineer — 5 months, fully online, with placement assistance. Week 1 is free.

Explore The AI-Era Agentic Engineer →