unreallabsai/unreal-agent is an async-first agent harness. It is a Go repository that splits the agent execution architecture into a library and a runner program that drives it, and the repository's latest release is v0.3.1. However, the provided records do not list what changed in v0.3.1.
The center of this repository is its structure that separates the orchestrator from the runner. The described components are input deduplication, session storage, the orchestrator, context assembly, the tool transformer, and the durable task runner. It converts tasks into a specification that can be serialized and hands them to the runner, and it manages the storage format and task versions directly. The design aims to make the boundaries of recovery and extension clear by allowing each component to be swapped. The license is MIT and the GitHub star count is 2209. The creation date is 2026-09-21.
The evidence stops here. The structure and design intent can be read, but the given records contain no benchmarks comparing performance, no real adoption cases, and no figures measuring the stability of async execution. The research record retains a 10-point scale with quality 8.0, momentum 7.6, novelty 7.0, and technology 7.0, but this is an evaluation left in the material, not independent verification. So it is still too early to decide on adoption on the basis of performance or reliability. The value visible now lies not in the breadth of the release changes but in the design that separates the orchestrator from the runner and makes the components interchangeable.