Unreal Labs has released Unreal Agent, a framework designed to reduce the cost of running AI agents in production. According to the company, Unreal Agent achieves up to 40% cost savings compared to Codex on production workloads and coding benchmarks, with no negative impact on performance.
The core innovation behind Unreal Agent is its approach to managing tool calls. Agents typically spend significant time and tokens managing tool interactions—waiting for results, polling for updates, and handling delays. Unreal Agent’s harness manages these tool calls completely asynchronously, removing the underlying model’s need to manage waits, polls, and heartbeats.
This architecture provides two main benefits. First, users can steer the agent without waiting for tool calls to finish. Second, the agent can schedule more tool work between model calls, improving cost efficiency. The framework achieves up to 40% savings versus Codex and up to 20% versus Pi across real workloads and agentic benchmarks.
Unreal Labs built the harness after identifying limitations in existing agent SDKs. CLI-oriented systems like Claude’s Agent SDK, according to the company, carry assumptions about local sessions and resource limits that don’t translate to production use. Supporting multiple providers adds compatibility complexity, with SDK upgrades potentially changing message formats and breaking tools. The company also noted that heavy dependency trees increase maintenance and supply-chain risk.
The cost savings come from two factors: minimal harness footprint with token-optimized tool results and simple prompts, and more tool work per model turn. Unreal Agent uses a straightforward asynchronous tool-calling model that allows issuing heavy tool calls without wasting tokens on polling.
Benchmark results show Unreal Agent performing competitively. On Terminal-Bench 4.0, it achieved 57.9% pass rate at $1,428 total cost versus Codex’s $2,350 at the same pass rate. On SWE-Atlas Codebase QnA, it scored 65.8% at $936 versus Codex’s 63.3% at $1,303. On DeepSWE 1.1, it achieved 72.4% at $1,367 versus Codex’s 69.0% at $1,633.
The Unreal Agent SDK currently offers a Go library for direct codebase integration, a runner executable, and a benchmark runner compatible with Harbor. The company notes that agent harness design is an emerging research area with many unexplored optimization opportunities.
Key facts
- Unreal Agent achieves up to 40% cost savings compared to Codex on production workloads and benchmarks
- The framework manages tool calls asynchronously, reducing token overhead and allowing users to steer agents without waiting for tool completion
- Users can schedule more useful tool work between model calls, driving improved cost efficiency
- Benchmarks show Unreal Agent outperforming Codex on multiple coding tasks while using fewer tokens and lower costs
- The SDK includes a Go library, runner executable, and Harbor-compatible benchmark runner