Something went wrong. Try again.
Something went wrong. Try again.
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123"""Letta Multi-Agent Financial Analysis Team Example================================================
This example demonstrates how to create a team of autonomous AI agents that work togetherto analyze financial markets, each with their own specialized knowledge and shared data.
Key Concepts:------------1. **Memory Blocks**: Letta agents use "memory blocks" to store and access information. Think of these as persistent knowledge bases that agents can read and write to.
2. **Shared vs Private Memory**: - Shared blocks (like market data) are attached to multiple agents - Private blocks (like proprietary models) are only attached to specific agents
3. **Agent Communication**: Agents can send messages to each other using tools, allowing them to share findings and coordinate analysis.
Architecture:------------- 3 Quant Agents: Each analyzes markets using different strategies - Momentum Quant: Uses price momentum signals - Value Quant: Uses fundamental value metrics - ML Quant: Uses machine learning predictions - 1 Portfolio Manager: Receives and consolidates findings from all quants
Memory Design:-------------- Global Shared Memory: "market-data" - All agents see the same market data- Local Private Memory: Each quant has their own model/strategy block- PM Aggregation Memory: "aggregated-signals" - Stores consolidated findings
Workflow:--------1. All agents read shared market data2. Each quant analyzes using their unique approach3. Quants send findings to the Portfolio Manager4. PM aggregates insights and produces final report"""
from letta_client import Lettaimport os
client = Letta(token=os.environ["LETTA_API_KEY"])project_id = "your-project-id"
# 1. Create shared memory blocks (global data)market_data = client.blocks.create( project_id=project_id, label="market-data", value="S&P500: 4800, VIX: 15.2, DXY: 102.5...", description="Shared financial market data")
# 2. Create individual quant memory blocksquant1_model = client.blocks.create( project_id=project_id, label="quant1-momentum-model", value="Momentum factor model: 12-month return signals...", description="Quant 1's proprietary momentum model")
quant2_model = client.blocks.create( project_id=project_id, label="quant2-value-model", value="Value factor model: P/E ratios, book values...", description="Quant 2's value investing model")
quant3_model = client.blocks.create( project_id=project_id, label="quant3-ml-model", value="LSTM predictions, feature importance weights...", description="Quant 3's ML model")
# 3. Create agents with mixed memory (shared + individual)quant1 = client.agents.create( project_id=project_id, name="quant-momentum", block_ids=[market_data.id, quant1_model.id] # Shared + Individual)
quant2 = client.agents.create( project_id=project_id, name="quant-value", block_ids=[market_data.id, quant2_model.id])
quant3 = client.agents.create( project_id=project_id, name="quant-ml", block_ids=[market_data.id, quant3_model.id])
# 4. Create portfolio manager with access to all findingspm_findings = client.blocks.create( project_id=project_id, label="aggregated-signals", value="", # Will be populated by quant reports description="Consolidated findings from all quants")
portfolio_manager = client.agents.create( project_id=project_id, name="portfolio-manager", block_ids=[market_data.id, pm_findings.id], tags=["pm"] # Tag for message routing)
# 5. Give quants ability to send findings to PMsend_to_pm_tool = client.tools.list(name="send_message_to_agents_matching_tags")[0]for quant in [quant1, quant2, quant3]: client.agents.tools.attach(agent_id=quant.id, tool_id=send_to_pm_tool.id)
# 6. Run analysisfor quant in [quant1, quant2, quant3]: response = client.agents.messages.create( agent_id=quant.id, messages=[{"role": "user", "content": "Analyze market and send findings to PM (tag: pm)"}] )