Hermes Framework for Agentic Research Released
A new open-source Python framework called Hermes has been released for building multi-agent AI systems for investment research. The MIT-licensed project uses Llama models to orchestrate specialized agents for tasks like data ingestion and signal validation, mirroring architectures used at large funds.
The Hermes framework's multi-agent approach mirrors a growing trend in quantitative finance to move beyond single, monolithic AI models. By assigning specialized roles—such as a Market Data Agent for real-time prices, a Sentiment Analysis Agent for news, and a Quantitative Analysis Agent for technicals—these systems can process diverse, unstructured data more effectively. This modular architecture enhances scalability and allows for more nuanced investment recommendations. Large language models are increasingly being used to translate unstructured financial information, like news and social media, into actionable trading signals. This emerging paradigm, sometimes called "the new quant," focuses on using LLMs for tasks like sentiment extraction, economic reasoning, and even generating auditable investment hypotheses. The goal is to create end-to-end decision systems that can reason over varied data types and translate that understanding into risk-controlled portfolios. A key challenge in agentic AI is "memory decay," where an agent's knowledge resets with each new session. The Hermes Agent, developed by Nous Research, tackles this with a multi-level memory system that creates "Skill Documents"—permanent records of successful complex task completions. This allows the agent to learn and recall successful workflows, improving its capabilities over time. Open-source backtesting libraries like Backtrader, vectorbt, and Zipline are essential tools for quantitative developers building on frameworks like Hermes. These Python-based engines allow for the rigorous testing of trading strategies against historical data, accounting for variables like transaction costs and slippage. This process is critical for validating the effectiveness of AI-generated signals before deploying them in live trading environments. The integration of specialized agents is showing significant efficiency gains in investment research, with some case studies reporting a 70-90% reduction in analyst research time. By automating the discovery, data collection, and analysis phases, these systems can evaluate hundreds of companies in parallel. A supervisor agent often manages the workflow, breaking down complex queries and delegating subtasks to the appropriate specialized agent. While many agentic systems are confined to specific platforms, some, like the Hermes Agent, are accessible through common messaging apps like Telegram, Discord, and Slack. This allows for a continuous feedback loop where a developer can initiate a task from a workstation and receive completion notifications or send follow-up instructions from a mobile device.
