Acorn: LLM Framework for Long-Running Agents with Structured I/O

Acorn is presented as an LLM agent framework designed to facilitate the creation of long-running agents with structured input/output (I/O). It emphasizes type safety through the use of Pydantic models for defining input and output schemas, and it automates tool calling, enabling agents to interact with external resources. The framework supports agentic loops, allowing for multi-turn execution and iterative problem-solving.

Key features of Acorn include automatic generation of tool schemas from type hints and docstrings, dynamic tool management (adding/removing tools during execution), parse error recovery through automatic retries, step callbacks for controlling loop behavior, and integration with LiteLLM, which provides access to various LLM providers. The framework also offers streaming responses and provider-level prompt caching to reduce latency and cost.

The framework's architecture is based on the concept of a `Module`, which serves as a base class for LLM agents. Developers configure the module with parameters such as the LLM to use, sampling temperature, maximum tokens, maximum loop iterations, input and output schemas, available tools, and caching settings. The framework also supports model fallbacks for automatic failover in case of provider issues.

Acorn aims to simplify the development of complex LLM agents by providing a structured and flexible framework. Its features, such as automatic tool schema generation and provider failover, can significantly reduce development effort and improve agent robustness. However, the framework's effectiveness depends on the quality of the underlying LLM and the design of the agentic loops.