2.2K ·GitHub
What is this?
Unreal Agent is an async-first agent harness written in Go — a toolkit for running LLM agents with persistent, forkable sessions, tool calling, and durable background work.
Unreal Agent is an async-first agent harness written in Go — a toolkit for running LLM agents with persistent, forkable sessions, tool calling, and durable background work.Is this a library I build on, or a ready-to-run app?It's a Go library: the harness/ package you build agents on. The repo also ships cmd/ executables that use the library and benchmarks/ runners.
It's a Go library: the harness/ package you build agents on. The repo also ships cmd/ executables that use the library and benchmarks/ runners.How do I get it running?No install steps are given. As a Go library, the first move is importing harness/ into your Go project; cmd/ contains executables showing how to use it in practice.
No install steps are given. As a Go library, the first move is importing harness/ into your Go project; cmd/ contains executables showing how to use it in practice.Do I need an account with a specific LLM provider to use it?No provider is named. The LLM Adapter component owns authentication and returns normalized responses, so you bring whichever provider implementation you plug in.
No provider is named. The LLM Adapter component owns authentication and returns normalized responses, so you bring whichever provider implementation you plug in.What happens if the process crashes mid-conversation?Sessions and operation state are persisted and support recovery, so a crash isn't fatal. But input deduplication is in-memory only, and an unsupported session version always errors explicitly on resume.
Sessions and operation state are persisted and support recovery, so a crash isn't fatal. But input deduplication is in-memory only, and an unsupported session version always errors explicitly on resume.What does 'async-first' mean in practice?Tool work isn't done inline: translators turn model requests into serializable operations that a separate actor runtime executes durably, while the coordinator keeps moving LLM turns forward.
Tool work isn't done inline: translators turn model requests into serializable operations that a separate actor runtime executes durably, while the coordinator keeps moving LLM turns forward.What tools does an agent get out of the box?A fixed registry of Bash, ViewImage, and a skill-use tool, each with a validating translator. The host selects which of them an agent session actually sees.
A fixed registry of Bash, ViewImage, and a skill-use tool, each with a validating translator. The host selects which of them an agent session actually sees.What exactly is a session?The agent's append-only, persisted history of a conversation. Sessions can be forked, and tool-call statuses are recorded atomically with the operations they produced.
The agent's append-only, persisted history of a conversation. Sessions can be forked, and tool-call statuses are recorded atomically with the operations they produced.Can tools execute somewhere else, like a remote sandbox?Yes — operations are versioned and always serializable, so a proxy operations manager can forward them to an operations manager running inside a remote sandbox and run the tools there.
Yes — operations are versioned and always serializable, so a proxy operations manager can forward them to an operations manager running inside a remote sandbox and run the tools there.What are similar projects?lightspeed, conductor, imp, OpenAgentCore, tardigrade, cayu, maka, dapr
lightspeed, conductor, imp, OpenAgentCore, tardigrade, cayu, maka, daprWhich topics is it in?actor-model, session-management, persistence, agent-runtime, orchestrator, event-sourcing, context-management, durable-execution, event-driven, tool-calling, agent-harness, concurrency
actor-model, session-management, persistence, agent-runtime, orchestrator, event-sourcing, context-management, durable-execution, event-driven, tool-calling, agent-harness, concurrency