We develop a vertically layered, protocol-oriented architecture for orchestrating autonomous FinAgents in a algorithmic trading task. It supports real-time decision-making by dynamically composing agents through task-specific execution graphs.
At the entry point, high-level strategic queries are issued via the Users Query interface. These are interpreted by the DAG Planner Agent, which constructs a directed acyclic graph (DAG) to encode task flow. The Orchestrator executes the DAG, coordinating communication and execution across agent pools.
The architecture includes the following specialized Agent Pools:
- Data Agent Pool: Acquires and preprocesses real-time or historical financial data
- Alpha Agent Pool: Generates predictive signals and quantitative factors
- Risk Agent Pool: Models portfolio exposure and applies constraints
- Transaction Cost Agent Pool: Estimates slippage and market impact
- Portfolio Construction Agent Pool: Allocates positions using alpha, risk, and cost inputs
- Execution Agent Pool: Routes and executes orders in external markets
- Attribution Agent Pool: Analyzes post-trade performance and contribution
- Backtest Agent Pool: Evaluates DAG performance over historical market data
All agent pools communicate with a centralized Memory Agent, which logs execution traces, model outputs, and evaluation results to support continual learning. Agent registration and health status are maintained by the Registration Bus.
The system uses four protocol layers to govern communication:
- MCP (Multi-agent Control Protocol): for task scheduling and DAG execution
- ACP (Agent Communication Protocol): for result reporting and synchronization
- A2A (Agent-to-Agent Protocol): for direct communication between dependent agents
- ANP (Agent Notification Protocol): for asynchronous, event-driven system alerts
The complete system design is illustrated below:
This framework enables composable, interpretable, and learning-augmented multi-agent orchestration, offering a flexible foundation for intelligent trading strategy research and deployment.
Four primary inter-agent communication protocols govern system operations:
| Protocol | Role |
|---|---|
MCP (Multi-agent Communication Protocol) |
Task scheduling and lifecycle control |
ACP (Agent Communication Protocol) |
Agent feedback and status reporting |
A2A (Agent-to-Agent Protocol) |
Peer-to-peer DAG-executed subtask coordination |
ANP (Agent Notification Protocol) |
Event-driven alerts and system-wide state propagation |
=======
├── docs/ Documentation and Sphinx sources
├── orchestrator/ DAG Controller, Orchestration engine, Bus, Protocols
├── agent_pools/ Modular agent pools (Alpha, Risk, Execution, etc.)
├── memory/ Memory Agent and DRL policy learner
├── config/ YAML-based system configuration
├── examples/ Strategy simulation and demo DAG runs
├── tests/ Unit and integration testing modules
├── Papers/ Whitepapers and system documentation
├── README.md Project overview and guide
├── requirements.txt Python dependency list
└── readthedocs.yml Build configuration for ReadTheDocs
- Research-grade multi-agent trading experiments
- Live or simulated trading strategy orchestration
- Benchmarking RL-based DAG optimizers in financial pipelines
- Self-adaptive strategy composition under varying market dynamics
Complete documentation is available in the docs/ directory and online at:
📘 https://finagent-orchestration.readthedocs.io
If you use this system or build upon it, please cite: