Agent Actors enables the creation of hierarchical, parallelized AI agent systems for complex task execution. It targets researchers and developers building autonomous AI workflows, offering a framework for agents to collaborate using a Plan-Do-Check-Adjust (PDCA) cycle and synthesized working memories.

How It Works

The system utilizes a tree structure where ParentAgent nodes plan and distribute tasks to ChildAgent nodes. These child agents execute tasks in parallel, performing "Do" and "Check" operations, and then report back. ParentAgents also handle "Adjust" phases, synthesizing insights from child agent outputs into a concise "working memory" for zero-shot LLM prompts. This approach facilitates complex problem-solving through distributed, iterative refinement.

Quick Start & Requirements

• Install via poetry: poetry add git+https://github.com/shaman-ai/agent-actors.git
• Alternatively, clone and install: git clone https://github.com/shaman-ai/agent-actors && cd agent-actors && poetry install --with dev --with typing
• Requires Python 3.10+.
• Demo walkthrough: https://www.loom.com/share/8e60585f069c4a9f8ac9f01204b41704
• Example usage and tests: test_system.py

Highlighted Details

• Implements Time Weighted Long-Term Memory using langchain.retrievers.TimeWeightedVectoreStoreRetriever.
• Features synthesized working memory for agents, limiting context to 1-12 items.
• Supports nested ParentAgent structures for complex agent hierarchies.
• Allows customization of vector stores, retrievers, and embedding functions.

Maintenance & Community

• Project is open for contributions, with specific requests for improved agent prompts, visualization tooling, evaluation data, and inter-agent communication.
• Links to community resources are not explicitly provided in the README.

Licensing & Compatibility

• Licensed under LGPLv3.
• Allows use within closed-source software, but modifications to the library must also be LGPLv3.

Limitations & Caveats

This is a proof of concept, not production-ready, and has primarily been tested with GPT-3.5. Running all tests may be time-consuming and hit API rate limits.