What is cayu?

Cayu is a Python runtime for building AI agents meant to run in production, not just as quick demos. The idea is that starting an agent prototype is easy, but the failures that matter show up later: a process crashes right after taking an action but before saving state, a model asks to use a tool it should not have access to, a run needs a person to approve something partway through, or the conversation grows too long for the model provider to accept another request. Cayu is built to handle those situations directly as part of the runtime, rather than leaving each application to solve them on its own. Instead of giving you a visual workflow builder or a simple chain-of-prompts helper, Cayu provides lower-level building blocks that your own application code combines: durable sessions that can survive a crash and be resumed, typed tools with clearly declared effects and policies around who can approve them, ways to manage a growing conversation so it does not overflow the model's limits, and structured events so a person can later reconstruct exactly what an agent run did. It also includes budget and cost tracking, support for multiple model providers such as OpenAI, Anthropic, Bedrock, and Vertex, and built-in ways to write tests and evaluations that check an agent's behavior, not just its final answer. To get started, you install Cayu with pip and use its command-line tool to scaffold a new project, which comes with one basic model-only agent, a test, and an evaluation already wired up. From there you build out your specific agent by adding tools and environment access as needed, following a documented loop of understanding the task, inspecting the project, planning a change, making it, then testing and evaluating the result. Cayu also ships an included FastAPI-based control plane and inspection dashboard for observing running agents. Cayu requires Python 3.11 or newer and separates concepts clearly: an agent describes who is acting, an environment describes what it can touch such as files or network access, and a session records one specific execution and its history. This separation is meant to keep production agent systems predictable even as they grow more complex.