Example: Chart Library research crew (researcher + analyst)

[grahammccain]

Adds a two-agent sequential Crew demonstrating the grounded-base-rates pattern for financial research.

The pattern

• Researcher — has tool access (get_cohort_distribution, explain_cohort_filters, refine_cohort_with_filters), no narrative license. Produces a numeric-only report.
• Analyst — has narrative license, no tool access. Turns numbers into a 150-word PM briefing.

The role separation enforces that every written claim maps back to a real retrieved number — a structural defense against hallucinated base rates in finance agents.

Tools

Wraps three HTTP endpoints from Chart Library:

• get_cohort_distribution(symbol, date, filters) → historical forward-return percentiles, MAE/MFE, realized vol, survivorship flag, named-event tags
• explain_cohort_filters(cohort_id) → ranks which filter would shift the distribution most
• refine_cohort_with_filters(cohort_id, extra_filters) → sub-second refinement of the stored cohort

Running

pip install crewai crewai-tools requests
export OPENAI_API_KEY=sk-...            # or swap for Anthropic
export CHART_LIBRARY_KEY=cl_...         # free at chartlibrary.io/developers
python cohort_research_crew.py NVDA 2024-06-18

About Chart Library

Open API serving 24M historical chart-pattern embeddings, 15K+ US equities, 10-year history. Free sandbox tier (200 calls/day). MCP server: pip install chartlibrary-mcp.

Draft for now — happy to adjust directory placement or style to match current crewAI-examples conventions.