Unreal Agent claims 40% cheaper coding runs via async tool handling

Unreal Labs has released Unreal Agent, an agent harness it says cuts operating costs by up to 40% versus OpenAI’s Codex and up to 20% versus Pi on production workloads and agentic benchmarks, with no measurable drop in task success. The core idea is to strip tool-call orchestration out of the model’s job. Instead of forcing the LLM to wait on, poll, and manage long-running tools, the harness logs each call as ‘in-progress’ the moment it fires, keeps executing it in the background, and only wakes the model when a result actually lands. That lets a single model turn kick off heavy, parallel work — spinning up a dev environment while simultaneously searching code and the web — without burning tokens on heartbeats and waits, and it lets users interject steering messages without stalling on unfinished calls.

The savings, the company argues, come from doing more tool work per model turn combined with a deliberately thin footprint: simple prompts, token-trimmed tool outputs, and no sub-agents or nested workflows. The pitch doubles as a critique of incumbent agent SDKs. Unreal Labs contends that CLI-oriented kits like Claude’s Agent SDK bake in assumptions about local sessions and subprocesses that don’t survive contact with production, that provider and version churn keeps breaking tool and message formats, and that heavy dependency trees add real supply-chain and patching burden. It also makes a security argument, favoring deterministic sandbox constraints — allow-listed hosts, scoped tokens, approval proxies — over harness hooks it views as harder to maintain and less robust.

Benchmarks were run on GPT-6 Astra at the xhigh setting against Codex and Pi across Terminal-Bench 4.0, SWE-Atlas, DeepSWE 1.1, and ALE-CLI, with pass-rate gaps described as benchmark noise rather than capability differences. The company ships a Go library, a Codex-style runner executable, and a Harbor-compatible benchmark runner, with code on GitHub. One honest caveat surfaces in the footnotes: the trick of holding both an in-progress and a final result for the same tool call is underspecified in the Responses API, and some non-OpenAI inference providers rejected the pattern during testing — a portability risk for anyone adopting the approach.