Unreal Agent

by Adi Gunawan 22 September 2026 5 min read 21x views

Quality versus disbursal frontier for coding agents, highlighting Unreal Agent alongside GPT-6 Astra at xhigh reasoning effort.

If you’re curious in frontier cost-efficiency for your AI agents, we’d affection to activity together! Get in touch: [email protected].

While deploying agents in the wild, we wanted them to react to users quickly and be cost-effective to run. We’ve noticed that agents expend a lot of period and tokens managing tool calls, which motivated us to build Unreal Agent alongside a harness that would decrease the model’s tool-management overhead.

The Unreal Agent harness manages tool calls in a entirely asynchronous way, relieving the underlying example of the need to oversee waits, polls, and heartbeats for tools.

This method drives two important benefits. First, it continually allows users to steer the delegate without the need to delay for tool calls to finish. Second, it allows the delegate to agenda additional helpful tool call activity between example calls, driving frontier disbursal efficiency. The current type achieves up to 40% disbursal funds compared to Codex and up to 20% compared to Pi in genuine workloads and on agentic benchmarks, which we portion here.

We accept harness scheme is a investigation area in its own right, alongside many promising ideas motionless to be researched and implemented.¹

Motivation and Architecture

If you try to build an agent-first product, you’ll quickly acknowledge that there’s no aureate way for implementing one. Big-brand vendors recommendation distinct SDKs to build agents, all alongside a distinct set of trade-offs that power not be immediately apparent.

At Unreal Labs, we have built a figure of agentic products and learned a few things concerning famous SDKs alongside the way.

For example, CLI-oriented SDKs specified as Claude’s Agent SDK transport assumptions concerning local sessions, subprocesses, and asset limits that don’t translate neatly into manufacturing use. Handling completion, cancellation, and backdrop tasks reliably frequently method construction your own lifecycle administration about them.

Supporting another providers adds compatibility work: switching API modes can interrupt tools or compaction, during SDK upgrades can alter communication formats and power integration rewrites. Heavy dependency trees add care and supply-chain hazard to a runtime we already need to comprehend and place ourselves.

Security and approvals that depend on harness hooks and specialized tools, in our experience, lean to necessitate additional care and be small sturdy than deterministic surroundings or sandbox constraints, exterior the harness: allowed/disallowed hosts, granular admission tokens, proxies alongside endorsement gates.

Along alongside these specialized motivations, we additionally wanted to build a harness that could continually obtain person steering messages without postpone and juggle heterogeneous tool calls without additional cognitive burden for the model. For example, we wanted the delegate to be capable to boot off a dev surroundings setup that power obtain minutes, during exploring the codebase and searching the web in parallel, all without additional token tax.

Every period Unreal Agent issues a tool call, we immediately append an event-log document that the tool has returned in the “in-progress” state, during continuing its implementation in the background. Once a tool really finishes, we append the outcome into the meeting log and call an LLM. Making this activity without breaking cache was an engaging engineering difficulty in itself.²

Cost Efficiency

On the surface, Unreal Agent achieves the identical outcomes alongside small example turns and small input tokens.

We trait disbursal funds to two factors:

Minimal harness footprint and careful engineering of tool output usage. Unreal Agent has uncomplicated prompts, token-optimized tool results, and no sub-agents or workflows.³
More tool activity per example turn. Unreal Agent has a straightforward asynchronous tool-calling example that is plainly explained to an LLM. This allows it to matter additional dense tool calls per example rotate without wasting tokens on polling or waiting for them.

Benchmarks

We’ve built Unreal Agent to provision genuine manufacturing workflows for us, but it looks fine in the benchmarks too. We tested it alongside GPT-6 Astra xhigh and compared it alongside Codex and Pi. Here are several of the results.

There are marginal differences in continue rate, which we trait to benchmark variance.

Terminal-Bench 4.0

GPT-6 Astra · xhigh. Codex (lb) is the leaderboard baseline; Unreal Agent and Pi runs are connected below.

Agent	Rate	Total $	In/trial	Out/trial	Turns	Tools	Harbor
unreal-agent	57.9%	1428	1.73M	32k	28	37	27133053
Codex (leaderboard)	57.9%	2350	—	—	—	—	—
Pi	55.0%	1827	2.83M	35k	44	57	6ccd097a

SWE-Atlas Codebase QnA

Agent	Rate	Total $	In/trial	Out/trial	Turns	Tools	Harbor
unreal-agent	65.8%	936	898k	15k	16	27	3d2fa057
Codex	63.3%	1303	1.69M	17k	22	21	11a440fb
Pi	64.0%	1033	1.29M	16k	24	60	da4ac972

DeepSWE 1.1

Agent	Rate	Total $	In/trial	Out/trial	Turns	Tools	Harbor
unreal-agent	72.4%	1367	1.60M	28k	26	38	2311ca63
Codex	69.0%	1633	2.19M	30k	30	29	e20ecafd
Pi	69.6%	1584	2.21M	30k	40	75	cd7d8de6

Agents’ Last Exam · ALE-CLI

Full continue rates and average scores are listed below. These runs are not on Harbor.

Agent	Full pass	Mean score	Total $	In/task	Out/task	Turns	Tools
unreal-agent	30.0%	59.7	217	0.76M	18k	18	23
Codex	29.0%	58.1	292	1.59M	15k	—	21
Pi	29.0%	59.2	262	1.19M	19k	27	37

We run mostly coding benchmarks since they are accessible on Harbor, which makes reproduction and verification easier, but the harness is domain-agnostic.