Unreal Agent

Unreal Agent is an asynchronous agent harness that removes tool-management overhead from the language model by handling tool calls outside the LLM loop: every tool call immediately appends an “in-progress” event to the session log, the tool runs in the background, and its final result is appended later to trigger the next model call. That design lets users steer agents without waiting, permits more heavy tool work per model turn, and reduces token taxes. Through a small prompt footprint, token-optimized tool outputs, and no sub-agents or workflows, Unreal Agent delivers up to 40% cost savings versus Codex and up to 20% versus Pi on production workloads and coding/science benchmarks while preserving or improving task pass rates.

Benchmarks show comparable or better performance: Terminal-Bench 4.0 yields the same pass rate as Codex with substantially lower cost; SWE-Atlas QnA and DeepSwe show higher pass rates (65.8% vs 63.3% and 72.4% vs 69.0%, respectively) with reduced dollars and fewer model turns. An engineering caveat: using two tool-call result items (in-progress + final) exposed inconsistencies across inference providers and an underspecified Responses API pattern that required careful handling. The SDK includes a Go library, a runner executable, and a Harbor-compatible benchmark runner, and the harness approach is presented as a distinct research direction for further efficiency gains.