The team behind an LLM video editor has released Gambit, an open-source agent harness for building reliable AI agents. It inverts traditional orchestration by making the LLM the primary driver, with compute tasks as secondary actions. Developers define agents in TypeScript or markdown 'decks' with typed interfaces, creating a modular system for composing complex workflows.

Gambit addresses common pain points in agent development, like brittle orchestration and difficult debugging. It provides built-in observability, a debug UI, and local tracing to make failures reproducible. The framework also includes automated 'grader' decks for evaluation and test agents for generating synthetic data, aiming to improve inference-time quality from the start.

Early design partners are already using Gambit to spin up bots in minutes, with tools like Codex or Claude Code. The goal is to enable truly open-source agents where prompts and logic are easily shared. By guaranteeing rubric-based grading, it aims to prevent issues like accidental PII leaks, making agent development more predictable and secure.