I am building Pleiad as an agent runtime that discovers capabilities from a user goal, generates task nodes, organizes dependencies into a graph, and executes the work.
I keep the model small: model, prompt, and tools. Tools and skills are selected from a capability catalog, while each task node keeps its execution boundary clear. The runtime exposes the same system through a terminal interface and an HTTP API.
I care about dynamic capability loading, structured results, observable execution, and a local-first runtime that can be inspected while it works.