Agents: Symbolic Learning Framework for Language Agents

Overview

Agents is an innovative framework for training language agents, drawing inspiration from the connectionist learning procedures used in neural networks. This systematic approach bridges the gap between traditional AI agent systems and neural network architectures, enabling more efficient and effective training of language-based AI agents.

Key Features

Analogous Structure to Neural Networks

• Agent Pipeline: Corresponds to the computational graph of a neural network
• Nodes: Equivalent to layers in a neural network
• Prompts and Tools: Act as the weights of a layer

Core Components

1. Loss Function: Implemented using prompt-based evaluation
2. Back-Propagation: Generates textual analyses and reflections for each node
3. Weight Optimizer: Updates symbolic components based on language gradients

Training Process

1. Forward Pass: Execute agent actions and store inputs, outputs, prompts, and tool usage
2. Loss Evaluation: Use prompt-based loss function to assess outcomes
3. Back-Propagation: Generate language gradients through textual analysis
4. Update: Modify symbolic components and computational graph structure

Multi-Agent Support

Agents naturally extends to multi-agent systems by:

• Treating nodes as different agents
• Allowing multiple agents to take actions within a single node

Use Cases and Demonstrations

The framework showcases its versatility through various expertly crafted scenarios:

1. NLP Classroom: Interactive communication between professor and students
2. Prisoner's Dilemma: Classic game theory scenario with rational agents
3. Software Design: Collaborative code generation with writer, tester, and reviewer
4. Database Administrator (DBA): System anomaly detection and diagnosis
5. Text Evaluation (ChatEval): Multi-agent referee team for assessing text quality
6. Pokemon: Interactive game world with multiple characters (available in release-0.1)

Benefits

• Enhances training efficiency for language-based AI agents
• Provides a structured approach to agent learning and optimization
• Supports both single-agent and multi-agent scenarios
• Enables complex, interactive simulations for various domains

Getting Started

To explore the capabilities of Agents, users can run the provided demo scenarios using the AgentVerse command-line interface. Each scenario offers unique insights into the framework's application in different contexts.

For developers and researchers, Agents opens up new possibilities in AI agent training, combining the strengths of symbolic AI with the learning capabilities of neural networks. This innovative approach has the potential to advance the field of AI agents, leading to more sophisticated and adaptable language-based AI systems.