Unreal Agent keeps reasoning while its tools are still running

Unreal Labs has open-sourced Unreal Agent, an agent harness built around asynchronous tool execution. Running GPT-6 Astra, the company reports the same Terminal-Bench 4.0 score as Codex at 39% lower cost.

Stop waiting for tools to finish

A build, test suite, or environment setup can keep an agent waiting for several minutes. Unreal Agent changes that loop by executing operations independently from model turns.

When a long-running command starts, the model can continue working, inspect other files, or initiate additional operations. Results are inserted back into the session as they become available.

The asynchronous architecture also allows multiple tool calls to be issued before the first one completes. Unreal Labs designed this to reduce model turns spent waiting or polling for command status, limiting the amount of context repeatedly sent back to the model. One Bash tool and less context

The savings do not come from switching models. Unreal Labs primarily attributes the difference to a deliberately small architecture: simple prompts, token-optimized tool results, no sub-agents, and more tool work completed per model turn.

The harness is organized around a coordinator, persistent session history, and an operation manager responsible for asynchronous execution. Sessions use append-only histories and can be resumed or forked.

Its tool registry is similarly compact, providing Bash, ViewImage, and Skills while allowing individual components to be replaced with alternative implementations.

Unreal Agent's source code includes the harness library, executables, and benchmark runners. Same score, 39% lower cost on Terminal-Bench

To isolate the harness as much as possible, Unreal Labs ran Unreal Agent with GPT-6 Astra at `xhigh` effort and compared it with Codex and Pi.

On Terminal-Bench 4.0, Unreal Agent reaches 57.9% at a reported total cost of $1,428. The Codex reference also reaches 57.9%, at $2,350. That difference produces the 39% cost reduction highlighted by Unreal Labs.

The same evaluation reports an average of 28 model turns and 37 tool calls for Unreal Agent, compared with 44 turns and 57 tool calls for Pi. Average input volume falls to 1.73 million tokens per trial for Unreal Agent versus 2.83 million for Pi.

Other company-published evaluations show a similar pattern. On DeepSWE 1.1, Unreal Agent scores 72.4% at $1,367, compared with Codex at 69.0% and $1,633. On SWE-Atlas Codebase QnA, it reaches 65.8% at $936, versus 63.3% and $1,303 for Codex. The harness becomes part of the cost equation

These results do not measure only GPT-6 Astra's ability to solve a task. They also illustrate the impact of the layer that organizes model calls, constructs context, and decides when tools should run.

Unreal Agent focuses on reducing that overhead rather than adding more orchestration. Its deliberate lack of sub-agents and complex workflows is therefore both an architectural choice and a limitation compared with more comprehensive agent environments.

Unreal Labs publishes execution traces and configurations alongside its benchmark results. The repository also includes Harbor-compatible benchmark tooling, allowing the reported runs to be inspected and reproduced rather than relying solely on the company's headline figures.