Unreal Labs Releases 'Unreal Agent' with Asynchronous Harness Design, Cutting Costs by Up to 40%

Unreal Labs has announced Unreal Agent, which applies a fully asynchronous harness to reduce tool call management overhead. By removing the burden of polling and heartbeats from the model, it enables user steering without waiting for tool call completion and allows scheduling more tool work between model calls.

Performance and Cost Efficiency

In real-world workloads and agentic benchmarks, it achieved cost reductions of up to 40% compared to Codex and up to 20% compared to Pi. This is attributed to a minimal harness footprint—including simple prompts, token-optimized tool results, and the absence of sub-agents/workflows—and the elimination of wasted waiting tokens.

Key benchmark results (based on GPT-6 Astra xhigh):

• Terminal-Bench 4.0: 57.9% ($1428) vs Codex 57.9% ($2350)
• SWE-Atlas Codebase QnA: 65.8% ($936) vs Codex 63.3% ($1303)
• DeepSWE 1.1: 72.4% ($1367) vs Codex 69.0% ($1633)

Technical Implementation and Security

Upon issuing a tool call, an event log with an 'in-progress' status is recorded immediately, and execution proceeds in the background. Upon completion, the result is appended to the session log before invoking the LLM. Cache maintenance was a key engineering challenge in this process. Additionally, it was noted that deterministic environment/sandbox constraints are more robust and require less maintenance than harness hook-based approvals.

SDK and API Findings

An SDK consisting of a Go library and a Runner executable is provided and has been published on GitHub. Notably, it was found that the handling of tool-call result items in the Responses API is poorly documented, causing rejections from some non-OpenAI inference providers, and explicit support for this pattern was proposed.