Skip to content
tama
Blog

Hands-on: 4 AI analysts researching a stock in parallel

tama team
tama team
Core contributors

Let’s build something real.

Four specialist AI analysts — fundamental, technical, sentiment, risk — run in parallel on any stock ticker. A portfolio manager reads all four reports and produces a final BUY / HOLD / SELL recommendation with price targets.

This is examples/18-stock-analysis from the tama repo, ported from CrewAI’s stock analysis tutorial. Same problem, different approach.

The architecture

Section titled “The architecture”
tamad "Should I invest in NVDA?"
      │
      ▼
  pipeline (FSM)
      │
      ├─► analysts (scatter → 4 parallel workers)
      │       ├─► analyst: FUNDAMENTAL
      │       ├─► analyst: TECHNICAL
      │       ├─► analyst: SENTIMENT
      │       └─► analyst: RISK
      │
      └─► portfolio-manager (synthesize → BUY/HOLD/SELL)

Three patterns working together: FSM connects the two phases, scatter fans out to the four analysts, each analyst runs a react loop with web search.

The files

Section titled “The files”

Five AGENT.md files. That’s the whole project.

agents/pipeline/AGENT.md

Section titled “agents/pipeline/AGENT.md”
---
name: pipeline
description: Stock analysis pipeline — 4 specialist analysts, then a portfolio manager.
version: 1.0.0
pattern: fsm
initial: analysts
states:
  analysts: portfolio-manager
  portfolio-manager: ~
---

This is the entry point. It’s pure routing — no prompt body needed.

agents/analysts/AGENT.md

Section titled “agents/analysts/AGENT.md”
---
name: analysts
description: Runs 4 specialist analysts in parallel for a given stock ticker.
version: 1.0.0
pattern: scatter
worker: analyst
call:
  model:
    role: thinker
  uses:
    - search-web
    - mem-set
---


You are a research director coordinating 4 specialist analysts.


Call start() to receive the stock ticker and question.


1. Extract the ticker symbol.
2. Store it: tama_mem_set("ticker", "<TICKER>")
3. Call parallel_run with 4 tasks:
   [
     "FUNDAMENTAL: <TICKER> — revenue, margins, P/E, cash flow, moat",
     "TECHNICAL: <TICKER> — trend, RSI, moving averages, support/resistance",
     "SENTIMENT: <TICKER> — news, analyst ratings, insider activity",
     "RISK: <TICKER> — macro, competition, regulation, valuation"
   ]
4. Store the combined reports: tama_mem_set("analyst_reports", "<all 4>")


Call finish(key="done", value="<brief synthesis>").

agents/analyst/AGENT.md

Section titled “agents/analyst/AGENT.md”
---
name: analyst
description: One specialist analyst — researches one dimension of a stock.
version: 1.0.0
pattern: react
call:
  model:
    role: thinker
  uses:
    - search-web
    - mem-append
---


You are a specialist financial analyst. Your task arrives prefixed with your type:
"FUNDAMENTAL: NVDA — ..."


Run 2–3 targeted searches for your specialty. Extract concrete data points.
Output in this format:


## <TYPE> Analysis: <TICKER>
**Verdict**: Bullish / Bearish / Neutral
**Key findings**: (3 bullet points with data)
**Summary**: (2–3 sentences)


Store your analysis: tama_mem_append("analyst_reports", "<your analysis>")
Call finish(key="done", value="<your formatted analysis>").

agents/portfolio-manager/AGENT.md

Section titled “agents/portfolio-manager/AGENT.md”
---
name: portfolio-manager
description: Synthesizes 4 analyst reports into a final investment recommendation.
version: 1.0.0
pattern: react
call:
  model:
    role: thinker
  uses:
    - mem-get
---


You are a senior portfolio manager. Read all 4 analyst reports and produce a
final recommendation.


1. tama_mem_get("ticker") and tama_mem_get("analyst_reports")
2. Weigh fundamental (30%), technical (25%), sentiment (20%), risk (25%)
3. Output:


# Stock Analysis: <TICKER>
## Recommendation: BUY / HOLD / SELL
**Conviction**: High / Medium / Low
## Key Drivers (bulls vs bears)
## Analyst Scorecard (table)
## Price Target & Entry
## Caveats


Call finish(key="done", value="<your complete report>").

Run it

Section titled “Run it”
Terminal window
git clone https://github.com/mlnja/tama
cd tama/examples/18-stock-analysis


export ANTHROPIC_API_KEY=sk-ant-...
export TAMA_MODEL_THINKER=anthropic:claude-sonnet-4-6


cargo run --bin tamad -- "Should I invest in NVDA?"

You’ll see 4 analysts spinning up in parallel, each running its own web search loop. A few minutes later, the portfolio manager synthesizes everything into a structured report.

What the output looks like

Section titled “What the output looks like”
# Stock Analysis: NVDA


## Recommendation: BUY
**Conviction**: High


## Key Drivers


### Supporting (Bulls)
- Dominant AI infrastructure position: 80%+ datacenter GPU market share
- Revenue up 122% YoY in latest quarter, margins expanding
- Technical: clean uptrend, RSI 58, holding above 200-day MA


### Against (Bears)
- Valuation stretched: P/E ~35x forward, premium to peers
- Export controls risk to China market (historically ~20% of revenue)
- Sentiment crowded: high short-term expectations


## Analyst Scorecard
| Dimension    | Signal   | Weight |
|-------------|----------|--------|
| Fundamental | Bullish  | 30%    |
| Technical   | Bullish  | 25%    |
| Sentiment   | Neutral  | 20%    |
| Risk        | Bearish  | 25%    |


## Price Target & Entry
- **Target**: $165 (+18% from current ~$140)
- **Entry**: Scale in on pullbacks toward $125–130
- **Stop-loss**: $118 (below key support)
- **Time horizon**: 12 months

What you’d change

Section titled “What you’d change”

Swap models. Use a faster/cheaper model for the 4 analysts and a stronger one for the portfolio manager — all without touching the agent files:

Terminal window
export TAMA_MODEL_THINKER=anthropic:claude-haiku-4-5
export TAMA_MODEL_THINKER_PM=anthropic:claude-opus-4-6

Or set it directly in tama.toml:

[models]
thinker = "anthropic:claude-haiku-4-5-20251001"

Add a 5th analyst. Create agents/macro-analyst/AGENT.md with pattern: react, add "MACRO: <TICKER> — interest rates, dollar strength, sector rotation" to the parallel_run call. Done.

Change the output format. Edit the portfolio manager’s prompt. The format is plain text in the AGENT.md body — it’s a diff in git.

Use a different ticker. tamad "What about MSFT?". The agents extract the ticker from the input.

vs the Python version

Section titled “vs the Python version”

The CrewAI version of this example is ~250 lines of Python across multiple files — agent class definitions, task objects, crew configuration, tool wrappers. If you want to change the output format, you edit a Python string embedded in a class method.

In tama: 5 Markdown files. Change the output format by editing the prompt. Add an analyst by adding a file. Diff your agent’s behavior in git.

The full example is at examples/18-stock-analysis. There are 22 more examples in the repo covering every pattern.

Tags: