Unreal Labs · 2026-09-22 · major
Unreal Agent — an open harness that runs tool calls in the background
Unreal Agent is an open-source agent harness from Unreal Labs that runs tool calls asynchronously, so the model never waits on a slow command. On DeepSWE 1.1 it scores 72.4% for $1,367, against Codex's 69.0% for $1,633.
An MIT-licensed Go harness that keeps the model working while its tools finish.
Key specs
| GitHub stars | 903 |
|---|---|
| Cost vs codex | 39% lower on Terminal-Bench 4.0 |
Quick facts
| Maker | Unreal Labs |
|---|---|
| License | MIT |
| Language | Go |
| Model providers | openai, openai-codex, openrouter, fireworks, ollama |
| Ships with | Go library, runner binary, benchmark runner |
| Repo | github.com/unreallabsai/unreal-agent |
Benchmarks
What is it?
Unreal Agent runs every tool call in the background instead of blocking the model on it. Unreal Labs published the Go harness on 22 September 2026 under the MIT license, along with a runner binary and a benchmark runner. It drives models from OpenAI, Codex, OpenRouter, Fireworks and Ollama.
How does it work?
A tool call is logged as in-progress the moment it starts, and the harness carries on with the turn. The result is appended once the work finishes and the model is called again, which Unreal Labs says leaves the prompt cache intact. Separate components handle the turn loop, the session history and durable execution of operations.
Why does it matter?
Cost is where the async design shows up. The same 57.9% pass rate as Codex on Terminal-Bench 4.0 costs $1,428 instead of $2,350 — about 39% less. Anyone running coding agents at scale pays for the turns a model burns polling a slow build or a long test run, and those turns go away.
Who is it for?
teams running coding agents at scale
Frequently asked questions
- How much cheaper is Unreal Agent than Codex?
- Unreal Agent's published runs cost 39% less than Codex on Terminal-Bench 4.0: $1,428 against $2,350 for the same 57.9% pass rate. On SWE-Atlas QnA it spends $936 against $1,303, and on DeepSWE 1.1 $1,367 against $1,633. Unreal Labs describes the saving as up to 40% on production workloads.
- Which models can Unreal Agent run?
- Unreal Agent picks a backend through the UNREAL_HARNESS_LLM_PROVIDER environment variable, which accepts openai, openai-codex, openrouter, fireworks and ollama. The model id is set separately with UNREAL_HARNESS_LLM_MODEL. The ollama option means the harness can drive a model running on your own machine instead of a hosted API.
- Is Unreal Agent open source?
- Unreal Agent is released under the MIT license, so commercial use and modification are both allowed. Unreal Labs publishes the Go library, the unreal-agent-runner executable and a Harbor-compatible benchmark runner in a single repository at github.com/unreallabsai/unreal-agent, which passed 900 stars within two days of the announcement.
- Who is behind Unreal Agent?
- Unreal Labs built Unreal Agent. The company says its team comes from CERN, Meta, Snap, Bloomberg and DeepMind, and that it is backed by Sequoia and First Round Capital. Unreal Labs states its interest is the harness, the runtime an agent acts in and the inference underneath, rather than the models themselves.
Try it
go install github.com/unreallabsai/unreal-agent/cmd/unreal-agent-runner@latest