Async-first agent harness
See the codeAn async-first agent harness from Unreal Labs.
| 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:
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.
Go
95.4%
Python
4.4%
Async-first agent harness
See the codeAn async-first agent harness from Unreal Labs.
| 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:
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.
Go
95.4%
Python
4.4%