AgentLoom
A YAML-configured runtime for building and operating multi-agent applications, run from a terminal Studio that shows evidence for every run, tracks a supervisor's goal to explicit completion, and gates memory promotion behind review.
AgentLoom treats a multi-agent system as an application with an execution contract: YAML defines a Supervisor, typed Workers, models, tools, Skills, hooks and runtime policy. Each Worker the Supervisor selects becomes a callable tool; simple Workers take an empty argument object and return text, while complex Workers declare JSON Schema input and output contracts that the runtime validates.
Every run gets an immutable ID, a manifest and versioned lifecycle events, with the same state readable from the terminal Studio, the CLI as JSON or JSONL, and a Python API; a preflight check can reject a run before storage is allocated. Goal Mode keeps one root Supervisor objective active across delegation and resumed sessions, and only that Supervisor can mark a goal complete with evidence. A separate self-learning system stores searchable history and evidence-gated memory, with promotion to project-wide scope always started by a person.
It requires Python 3.12 or newer.