Show HN: Jeju – a local-first agent harness with inspectable runs (github.com)

Jeju, an innovative local-first agent harness, has been announced as an experimental tool aimed at developers who prioritize explicit and inspectable agent behavior. By defining agent characteristics in a streamlined configuration manifest—including model provider, instructions, runtime loop, and evaluation criteria—Jeju allows users to run agents in a controlled local environment. Each execution logs meaningful outcomes in a trajectory file, enabling subsequent inspection and evaluation to enhance the agent's performance through its unique "jeju evolve" feature. Significantly, Jeju empowers developers to create focused, reusable agent bundles tailored to specific workflows like bug fixes, as showcased in a Python application example. The harness captures the lifecycle of agent executions, allowing for the analysis of decisions and impacts within set limits. This structure not only promotes standardization and repeatability but also emphasizes the importance of evaluation-guided improvements. With configurations kept as the single source of truth, Jeju represents a shift towards more modular, auditable, and efficient AI/ML agent development, paving the way for a deeper understanding of agent workflows and behaviors.