AgentShip

AgentShip addresses the gap between an agent that works in a notebook and one that is actually shippable: a REST API, session storage, cost-tracking tracing, streaming with its error cases, and packaging, which is otherwise a few thousand lines of infrastructure unrelated to what the agent itself does. Rather than reimplementing that infrastructure, it integrates existing libraries behind small, stable interfaces, so replacing one of them, such as swapping out LangGraph, is a matter of changing one adapter package rather than rewriting the application.

An agent or a team of agents is authored in YAML or Python, and a coordinator can route work to team members based on each member's own description, with a member defined inline, referenced from its own YAML file, or reached remotely over the A2A protocol. A built-in offline engine that needs no model provider or API key lets the transports, authentication, error handling, and studio interface be explored without cost, and the same spec can be run as a one-off command or served as a REST, SSE, and WebSocket API with a studio UI.