Unreal Agent

Unreal Agent is an open-source agent harness from Unreal Labs, written in Go and released under the MIT License. It provides a library for running LLM agents asynchronously, plus executables that use it: a cloud agent runner and a terminal UI. The latest release listed is v0.3.1.

What It Is

Unreal Agent is a harness library for building and running agents. The repository contains the harness library, the executables unreal-agent-runner (cloud agent runner) and unreal-agent-tui (a terminal UI), and benchmark runners.

How the Harness Is Structured

The design is built around sessions, inputs and operations. Inputs carry caller-supplied globally unique IDs, and a session inbox deduplicates them. Sessions are append-only persisted histories that can be forked.

A coordinator persists accepted inputs, runs LLM turns, resolves tool translators through a registry, and dispatches committed operations. Tool translators validate model tool calls and turn them into serializable operations, without performing I/O. The tool registry owns fixed Bash, ViewImage and skill-use definitions. An operation manager runs durable operations as an actor runtime, and the local implementation is swappable.

Other components are a session store (recovery and forks), a context builder that assembles model input in memory and records anything omitted, truncated or compacted, and an LLM adapter that handles provider authentication, cancellation and errors.

Extensibility and Compatibility

Components are composable, and alternative implementations of their interfaces are encouraged. For example, a proxy operations manager could send serialized operations to a local operations manager inside a remote sandbox so tools execute there. The project intends to preserve these invariants: session-store items are serializable with a versioned storage format, operations are versioned and serializable, and an unsupported session version causes an explicit error on resume.