unreallabsai/unreal-agent
Async-first agent harness
What it solves
Unreal Agent provides an asynchronous-first agent harness designed for reliability and durability. It addresses the challenge of managing complex agent workflows by providing a structured framework for handling inputs, persisting session history, and executing tools through a decoupled architecture that separates tool translation from actual execution.
How it works
The system is built around a coordinator that manages the lifecycle of LLM turns and tool usage. It uses a session inbox for input deduplication and a session store to persist history and operation states, allowing for recovery and session forking. The process follows a specific flow: a context builder assembles the model input, an LLM adapter sends it to a provider, and tool translators validate model requests and turn them into serializable operations. These operations are then handled by an operation manager that executes the work asynchronously.
Who it’s for
Developers building AI agents that require durable execution, session persistence, and a clear separation between the tool definition and the execution environment (such as running tools in remote sandboxes).
Highlights
- Async-first architecture: Designed for non-blocking execution of agent tasks.
- Durable Sessions: Supports append-only persisted history with the ability to fork sessions.
- Decoupled Tooling: Separates tool translation (synchronous validation) from operation execution (asynchronous work).
- Composable Components: Provides swappable interfaces for the session store, LLM adapter, and operation manager.
Related
- Project
- Project
- Project
- Dispatch
- Project