An async-first agent harness from Unreal Labs.
- harness/ — the library.
- cmd/ — executables that use the library.
- 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.
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.