Unreal Agent ships as an open, async-first agent harness

Unreal Labs shipped Unreal Agent

Unreal Labs published Unreal Agent on 22 September 2026: an open source, MIT-licensed agent harness written in Go that runs a model's tool calls asynchronously instead of making the model wait for each one to finish. The company says it measured up to 40% lower cost than OpenAI's Codex and up to 20% lower cost than Pi on production workloads and coding benchmarks, with no drop in task success. The code is on GitHub at unreallabsai/unreal-agent.

Unreal Agent is a harness: the layer that sits between a model and its tools, deciding how tool calls get scheduled, retried and fed back into the conversation. It handles a tool call by logging it as in progress and letting execution continue in the background, so a user can send a new message or the agent can start another tool call without waiting for the first one to return, then it appends the real result and calls the model again once the tool finishes.

What changed

• Every tool call gets an immediate in-progress record in the session's append-only history. The tool keeps running; when it returns, the result is appended and the harness calls the model again. Nothing about the pattern needed a new API, but Unreal Labs notes in the post that the two-tool-call-result-per-turn pattern (one in-progress, one final) is underspecified in OpenAI's Responses API and caused rejections with some non-OpenAI inference providers.
• The harness lets a user steer the agent mid-task without waiting for a running tool call, and lets the agent schedule more tool work per model turn instead of spending tokens on polling or heartbeats.
• Unreal Labs says the design deliberately avoids CLI-oriented assumptions it found in SDKs like Claude's Agent SDK, such as local sessions and subprocess lifecycles that do not translate cleanly to a production service, and it keeps security and approvals as deterministic constraints outside the harness (allowed and disallowed hosts, scoped tokens, proxies with approval gates) rather than as harness hooks.
• The SDK ships as a Go library for embedding, a runner executable comparable to claude -p or codex exec, and a benchmark runner compatible with the Harbor framework.
• On benchmarks run with GPT-6 Astra at xhigh reasoning effort, Unreal Agent's own numbers show: Terminal-Bench 4.0 at 57.9% pass for $1,428 total (28 turns, 37 tool calls) against Codex's leaderboard baseline of 57.9% for $2,350 and Pi's 55.0% for $1,827 (44 turns, 57 tools); SWE-Atlas Codebase QnA at 65.8% for $936 against Codex's 63.3% for $1,303 and Pi's 64.0% for $1,033; DeepSWE 1.1 at 72.4% for $1,367 against Codex's 69.0% for $1,633 and Pi's 69.6% for $1,584; and on Agents' Last Exam (ALE-CLI), a 30.0% full-pass rate for $217 against Codex's 29.0% for $292 and Pi's 29.0% for $262.

Why it matters for agent builders

Most teams building an agent inherit whatever tool-calling loop their SDK gives them, and that loop is usually synchronous: the model waits for a tool, gets the result, then acts. Unreal Agent's numbers argue that the wait itself is a cost center, not just a latency annoyance, because a model that has to poll or re-explain state to itself burns tokens doing it. That claim lines up with the University of California, Berkeley Sky Lab team's HarnessTax study, which Unreal Labs cites and which found that swapping a coding agent's harness can move cost by up to 5x while barely moving success rate. If the harness is where a large share of an agent's operating cost actually lives, the harness is a decision worth making deliberately rather than accepting the vendor default.

What to watch

Unreal Labs flags its own edge case: the in-progress-plus-final tool result pattern isn't formally specified in OpenAI's Responses API, and some non-OpenAI inference providers rejected it during testing. That is a compatibility risk for anyone who wants to run Unreal Agent against a model provider Unreal Labs has not already validated. The Hacker News discussion also raised a separate, harder problem the harness does not solve on its own: context bloat from wiring in many MCP servers or tools at once, with commenters pointing to code-mode execution (a single execute-code tool instead of many discrete ones) and skill-tree style tool discovery as alternatives worth watching alongside async scheduling.

Key takeaways

• Unreal Agent is an open source, MIT-licensed, Go-based agent harness from Unreal Labs, released 22 September 2026, that runs tool calls asynchronously instead of blocking the model on each one.
• Unreal Labs reports up to 40% lower cost than Codex and up to 20% lower cost than Pi across four benchmarks run with GPT-6 Astra at xhigh, with pass rates matching or slightly exceeding both.
• The harness deliberately keeps security and approval logic outside itself, as sandbox and network constraints, rather than as harness-level hooks.
• Unreal Labs' own post flags that the async tool-result pattern it relies on is underspecified in OpenAI's Responses API and failed on some non-OpenAI providers during testing.
• The release cites UC Berkeley's HarnessTax study, which found harness choice alone can move a coding agent's cost by up to 5x without moving its success rate.

FAQ

What is an agent harness?

An agent harness is the software layer between a language model and its tools that decides how tool calls are scheduled, retried, logged and fed back into the conversation. It sits below the SDK a developer writes against and above the raw model API.

How does Unreal Agent save cost compared to Codex and Pi?

It handles tool calls asynchronously: it logs a call as in-progress and keeps it running in the background instead of pausing the model, which lets it pack more tool work into fewer model turns and avoid spending tokens on polling or waiting. Unreal Labs measured this producing up to 40% lower cost than Codex and up to 20% lower cost than Pi on its benchmark set.

Is Unreal Agent open source?

Yes. It is published under the MIT license at github.com/unreallabsai/unreal-agent, with a Go library, a runner executable and a Harbor-compatible benchmark runner.

Source: Unreal Agent Discussion: Hacker News

Anyone comparing harnesses before wiring one into a production agent can browse the current field in AI Agents Listing's agent harness listings.