Traders Agents
Ed DonnerFour peer Trader agents on a shared scheduler. Each loads a strategy prompt, uses MCP tools for market and research, then decides trade or hold and may notify.
Scheduler · OpenAI agents · MCP accounts · market · research · notifications
Overview
Layers 2–5 on the Mental Model. Four Trader agents (Warren Patience, George Bold, Ray Systematic, Cathie Crypto) share the floor timer. Each builds an OpenAI Agents SDK agent with MCP for accounts, push, and market, plus a nested Researcher tool. Env can assign different models per trader (USE_MANY_MODELS) or one shared gpt-4o-mini.
Behavior
- Scheduled peers (Parallelization). The floor wakes every trader on an interval; each decides trade or rebalance alone. See Building effective agents.
- Strategy as prompt. Persona and style live in
trader_instructions; message istrade_messageorrebalance_messagebased ondo_trade, which flips after each run. - Orchestrator-Worker. Nested Researcher has its own MCP belt (Brave, Fetch,
memory/{name}.db); the trader places orders and push. - Tools over MCP. Trader always attaches accounts, push, and market (local
market_server.pyor Polygon MCP). - Max turns and traces.
Runner.runwithmax_turns=30;LogTracerattributes rows per trader for the dashboard.
Tick flow
One scheduled tick for a single trader. Diagram key: ✦ AI agent · second line = Anthropic workflow · purple = actor · white = service.
---
config:
theme: default
themeVariables:
background: "#ffffff"
fontSize: "14px"
---
flowchart TB
sched[Scheduler]
trader["✦ Trader agent
Parallelization"]
strat[Load strategy and account]
research["✦ Researcher tool
Orchestrator-Worker"]
market[Market MCP]
decide[Trade or rebalance]
acct[Accounts MCP]
push[Push MCP]
board[Dashboard logs]
sched -->|wake| trader
trader --> strat
strat --> research
strat --> market
research --> decide
market --> decide
decide --> acct
decide --> push
push --> board
classDef agent fill:#e8f0ee,stroke:#0f3d36,stroke-width:2px,color:#0d0d0d
classDef service fill:#fafaf8,stroke:#c8c8c2,stroke-width:1px,color:#0d0d0d
classDef actor fill:#f3eef8,stroke:#6b4f9a,stroke-width:2px,color:#0d0d0d
class trader,research agent
class sched,strat,market,decide,acct,push,board service
Entry: trading_floor.py → traders.py Trader.run(). Researcher MCP never attaches to the trader process belt.
Traders are peers on a timer, not a supervisor chain. Layer map on Mental Model. Part of Project.