Unreal Labs released an open-source agent harness that runs tool calls asynchronously and claims lower costs than Codex on selected benchmarks.

Unreal Labs has published Unreal Agent, a coding and general-purpose agent harness designed to issue independent tool calls concurrently. The company argues that many agent loops wait unnecessarily for one command to finish before starting another. By identifying operations that do not depend on each other, the harness can reduce idle time and sometimes total token use.

Why it matters

Model choice is only one part of agent performance. The harness decides what context the model receives, which tools it can use, how results return and when a failed step is retried. Two systems using the same model can therefore differ substantially in speed, cost and success rate.

Asynchronous execution is familiar in ordinary software. An agent may be able to search several directories, inspect independent files or run unrelated tests at the same time. Serial execution makes the model wait for each result. Parallel execution can reduce wall-clock time, though it may waste work if later decisions show that some calls were unnecessary.

Unreal Labs reports that its harness cuts costs by as much as 40% compared with Codex without reducing benchmark performance. Its published examples include the same 57.9 score on Terminal-Bench at a reported cost of $1,428 versus $2,350, and a 65.8 result on SWE-Atlas compared with 63.3. These are creator-run comparisons, not independent audits.

Benchmark interpretation depends on configuration. A harness can change prompts, tool interfaces, timeouts, retry policy and model effort. Cost also varies with provider pricing and caching. A fair comparison needs identical task sets, model versions and acceptance criteria, plus disclosure of failed runs. Headline totals without that context can hide important trade-offs.

Concurrency introduces its own engineering problems. Two tools may modify the same file, consume rate limits or rely on shared state. The harness must know which calls are safe to overlap and how to cancel or reconcile work. A faster agent is not better if races produce inconsistent patches.

Open-source availability helps scrutiny. Developers can inspect the orchestration logic, reproduce published runs and adapt controls to their environment. It does not automatically verify the benchmark claims; reproducibility still requires datasets, exact configurations and sufficient compute.

The release arrives as model providers cut token prices, so savings can compound. A cheaper model inside a more efficient harness may reduce the cost of a completed task more than either change alone. Conversely, an aggressive harness can spend more by launching too much speculative work.

Teams evaluating Unreal Agent should use real repositories and track accepted outcomes, wall-clock time, model tokens, tool calls and human corrections. They should also stress-test file conflicts and cancellation. The relevant question is not whether concurrency sounds efficient, but whether it produces correct work more cheaply under controlled conditions.

Verification

  1. VERIFIED — Unreal Labs released Unreal Agent and published its source. Primary source: https://unreallabs.ai/blog/unreal-agent/
  2. VERIFIED — The harness supports asynchronous tool calls. Primary source: https://unreallabs.ai/blog/unreal-agent/
  3. VERIFIED AS VENDOR-REPORTED RESULTS — The cost and benchmark comparisons were run or published by Unreal Labs. Primary source: https://unreallabs.ai/blog/unreal-agent/
  4. UNVERIFIED — Independent reproduction of the claimed 40% saving was not located.

Glossary candidates

  • Agent harness: Software that manages a model’s tools, context and execution loop.
  • Asynchronous: Allowing work to proceed without waiting for each earlier operation to finish.
  • Race condition: Incorrect behavior caused by operations interacting in an unexpected order.

Cold-reader sentence: Unreal Agent uses concurrent tool calls to pursue cheaper coding work, but its reported benchmark savings still need independent reproduction.