Result of program evaluation containing both metrics and optional artifacts This maintains backward compatibility with the existing dict[str

Represents a program in the database

Result that can be pickled and sent between processes

Configuration for a single LLM model within an ensemble.

Manages an ensemble of Language Models, allowing for weighted sampling and parallel generation.

Generates prompts for code evolution

Master configuration for LongHorizon

Encapsulates logic for autonomous web research using the Browser Use library.

Configuration for OpenAI Computer Use (Visual Evaluator)

Evaluates programs and assigns scores The evaluator is responsible for executing programs, measuring their performance, and assigning sc

Represents a single evolution trace entry

Main controller for LongHorizon Orchestrates the evolution process, coordinating between the prompt sampler, LLM ensemble, evaluator, an

Configuration for Browser Use (Research Agent)

Configuration for the program database

A client class to handle requests to embedding models from OpenAI or Azure.

Configuration for program evaluation

A dataclass to hold the final result of an evolution run.

Manages evolution trace logging with support for multiple formats

Configuration for the overall LLM setup, including ensembles.

LLM interface using OpenAI-compatible APIs

Controller for process-based parallel evolution

Database for storing and sampling programs during evolution The database implements a combination of MAP-Elites algorithm and island-bas

Configuration for prompt generation

Resulting program and metrics from an iteration of LongHorizon

A simple task pool for managing and limiting concurrent asyncio tasks.

Manages templates with cascading override support

Manages the process of running a program that produces visual output, capturing the screen, and using a multimodal LLM to evaluate the result

Configuration for evolution trace logging

Abstract base class for LLM interfaces