Go harness for LLM agents that forks sessions and sandboxes tools

What it does Unreal Agent is a Go library for building LLM agents that must survive crashes, redelivered events, and long-running tool calls. It gives you a coordinator to drive LLM turns, a session store for append-only persisted history, and a tool registry where validation happens synchronously on the event loop while actual work ships out as async, serializable operations.

The interesting bit The design enforces a strict no-I/O rule inside the coordinator: tool translators produce operations but cannot suspend the loop or touch the network. Because operations are just serialized descriptions of work, you can slot in a proxy manager to run them inside a remote sandbox, keeping the coordinator’s world small and deterministic.

Key highlights

• Session history is append-only, versioned, and forkable; resuming an unsupported version always fails explicitly.
• The context builder assembles prompts in memory with no persistence dependencies and reports exactly what got truncated or omitted.
• Inputs carry caller-supplied globally unique IDs, and the inbox deduplicates them across redeliveries.
• The LLM adapter normalizes provider responses and owns authentication, cancellation, and errors.
• The operation manager is swappable, so durable tool execution can live in another process or sandbox.

Verdict Pick this up if you’re building production Go agents that need durable, auditable sessions and clean separation between planning and execution. Look elsewhere if you want a quick Python script or a fully-baked chatbot framework.