Unreal Labs' Unreal Agent Cuts AI Coding Costs 39% Without Losing Performance
Unreal Labs open-sourced a Go-based async agent harness that reportedly runs Terminal-Bench 4.0 at 39 percent lower cost than Codex plus Astra.
- Unreal Labs open-sourced Unreal Agent, an async-first agent harness written in Go under MIT license.
- Claims 39 percent lower cost than Codex plus Astra on Terminal-Bench 4.0 with matching pass rate.
- Tools run independently of model turns, so the LLM avoids spending tokens polling or waiting.
- Design uses a single bash tool, minimal prompts, no sub-agents, and encourages batched calls.
- Sessions are append-only, forkable, versioned, and support crash recovery via a swappable operation manager.
- Design details in the Unreal Labs blog post covering benchmarks against Codex and Pi.
Unreal Labs has released Unreal Agent, a general-purpose agent harness written in Go and distributed under the MIT license. The company claims the harness can cut agent costs by up to 40 percent without reducing task performance. Its headline result is a 39 percent cost reduction against Codex plus Astra on Terminal-Bench 4.0, with a roughly equivalent pass rate.
Unreal Agent changes the orchestration layer around a language model. An agent harness manages prompts, tool calls, execution state, and results; it does not supply a new model. Unreal Labs attributes the reported savings to asynchronous tool execution, compact prompts, token-efficient outputs, and a deliberately small tool surface. The source, setup instructions, and examples are available in the GitHub repository, alongside a companion design post.
| Property | Details |
|---|---|
| Language | Go |
| License | MIT |
| Core tools | Bash and an image viewer |
| Execution model | Asynchronous, durable operations |
| Session model | Append-only, recoverable, and forkable |
Async execution without polling
Long-running tools execute as asynchronous operations, which lets the model submit several independent calls during one turn. The coordinator tracks those operations and consumes their results as they complete. Users can also send new instructions while tools remain active, avoiding a forced wait for every command to finish.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
