Unreal Agent An async-first agent harness

Unreal Labs released Unreal Agent, an async-first agent harness organized into a harness library, cmd executables, and benchmark runners. The harness defines an input/inbox/session model with a coordinator, session store, context builder, LLM adapter, tool registry, tool translator, and operation manager, and it commits to versioned, serializable session-store items and operations with explicit errors on unsupported session versions during resume. Tool translators run synchronously on the coordinator's event loop and must not perform I/O, while operations are dispatched for asynchronous execution and can be proxied to a remote sandbox.

An async-first agent harness from Unreal Labs. - harness/ https://github.com/unreallabsai/unreal-agent/blob/main/harness — the library. - cmd/ https://github.com/unreallabsai/unreal-agent/blob/main/cmd — executables that use the library. - benchmarks/ https://github.com/unreallabsai/unreal-agent/blob/main/benchmarks — benchmark runners. - Input : an event with a caller-supplied globally unique ID that remains stable across redeliveries. - Inbox : session-scoped, in-memory deduplication of external, control, and crash inputs. - Session : append-only persisted history that can be forked. - LLM turn : the coordinator-managed sequence around one logical LLM request. - Tool : a capability described by a schema and bound to a translator. - Tool call : a model-produced request to use a tool. - Tool translator : validates a tool call and translates it into one or more operations. It runs synchronously on the coordinator's event loop and must not perform I/O or suspend the loop. - Tool call status : the translation outcome: a validation error or references to submitted operations. Operation execution state is tracked separately; the translator formats these into a model-facing result. - Operation : a serializable description of work produced by a tool translator for asynchronous execution. Implementations are encouraged to use the available primitives https://github.com/unreallabsai/unreal-agent/blob/main/harness/primitives . | Component | Responsibility | |---|---| | Session inbox | Volatile, session-scoped input idempotency. | | Coordinator | Persist accepted inputs, run LLM turns, resolve tool translators through the registry, and dispatch committed operations. | | Session store | Persist canonical session history and operation state; support recovery and forks; atomically record tool-call status with operations. | | Context builder | Statefully assemble model input in memory. Return the model input together with a record of anything omitted, truncated, or compacted. Perform no I/O and accept no persistence dependencies. | | LLM Adapter | Send prepared model input to a provider and return a normalized completed response. Own authentication, cancellation, and provider errors. | | Tool registry | Own the fixed Bash, ViewImage, and skill-use definitions and their translators; expose the host-selected set. | | Tool translator | Validate a tool call and produce its status and operations. Format a recorded call status and prepared operation output into model results. Perform no I/O. | | Operation manager | Actor runtime for durable operations. The local implementation is swappable. | Harness components are composable, and alternative implementations of their interfaces are encouraged. We intend to preserve these invariants: - Session-store items are serializable, and the storage format is versioned. - We'll do our best to maintain backwards compatibility for sessions. An unsupported session version will always cause an explicit error on resume. - Operations are versioned and always serializable. For example, a proxy operations manager can send serialized operations to a local operations manager running in a process inside a remote sandbox, allowing tools to execute there.