The prevailing approach to computer-use agents couples a model with a domain-specific harness: a browser or desktop environment equipped with human engineered tools that are fixed before task execution. As models' coding capabilities improve, the GUI native and static harness prevents them from direct programmatic operation on system state, as well as flexible construction of tools. To this end, we introduce CUAWright, a minimal terminal harness of roughly 3K lines of code that uses bash commands as its sole action interface, and a file system as its evolvable space for dynamically creating tools and managing the context. We conduct comprehensive experiments across a wide range of digital tasks, and demonstrate that by giving the agent a minimal, programmable interface, it achieves substantially stronger results compared to their GUI or hybrid CLI interface across a wide range of tasks. On OSWorld 2.0, CUAWright delivers a 33.2% relative improvement in partial reward while reducing estimated cost by 37.5% compared with the published GPT-5.5 baseline. On Online-Mind2Web and the long horizon Odysseys benchmark, CUAWright substantially outperforms GUI native harness by 4.7% and 44.0% in success rate, respectively. Furthermore, we found the gains extend to CAD applications that require accurate visual understanding and CLI interaction: on CADGenBench and BenchCAD, our unified harness yields 8.1%-41.6% relative improvements over other CLI based harnesses with GPT-5.5. Together, these results suggest digital environments are far more programmable than their GUI interfaces imply, and a minimal terminal-focused harness is the key for better performance and efficiency.
Figures & tables
Figure 1: CUAWright demonstrates strong performance across different digital task benchmarks. For each benchmark, we compare CUAWright with the existing GUI or hybrid CLI based approach with highest performance score using the same underlying model. CUAWright demonstrated substantially better performance compared to the baselines.
Figure 2: The CUAWright architecture. Terminal serves as the sole interface between the agent and the task environment. At each step of task execution process, the agent model receives command line output as observation or call a CLI tool to load image into context on demand, and output commands that gets executed. The underlying execution environment contains a runtime, an expandable set of installed CLI tools and applications, and a file system as the agent’s workspace to programmatically manage system state and craft new tools.
Online Mind2Web
Odyssey
Easy
Medium
Hard
Overall (%) ↑
Easy
Medium
Hard
Overall (%) ↑
GUI baseline GPT-5.4
89.9
81.0
76.6
83.4
80.0
58.7
3.7
33.5
CUAWright GPT-5.4
90.3
86.5
88.8
88.1 [ Δ +4.7]
95.6
78.3
69.7
77.5 [ Δ +44.0]
CUAWright GPT-5.5
92.9
87.0
90.5
89.4 [ Δ +6.0]
93.3
89.1
85.3
88.0 [ Δ +54.5]
Table 1: Performance comparison on two web agent benchmarks: Online Mind2Web and Odyssey.
Model / harness
Partial (%) ↑
Strict (%) ↑
Cost / task (USD) ↓
Published price–performance baselines
GPT-5.6 Luna Max
45.57
–
$7.79
Claude Opus 4.8 Max
54.80
20.6
$31.4
Claude Opus 5 Max
68.3
31.4
$28.4
GPT-5.5 xhigh
GPT-5.5, official
47.5
13.9
$16.1
Table 2: OSWorld 2.0 desktop results and price–performance comparison. Scores are mean partial reward; higher is better. Costs are API cost per task; lower is better.
Figure 3: The price--performance comparison of CUAWright and the official GUI-native harness on OSWorld 2.0. CUAWright costs are computed from token usage cost per trajectory using public prices as of September 2026 1 1 1 OpenAI introduced promotional pricing for GPT-5.6 Sol on August 21, 2026, reducing the API price from 5/30 to 4/20 per million input/output tokens. Using the pre-promotion pricing, CUAWright with GPT-5.6 Sol costs 5.7pertask,stillsubstantiallybelowtheleaderboardreportedcostof13.9 per task. , while external baselines use their reported costs.
Harness
Agent
PassRate (%) ↑
Overall ↑
Codex CLI
GPT-5.5
35.1
49.9
OpenClaw
GPT-5.5
33.3
46.6
Hermes
GPT-5.5
31.6
46.6
CUAWright
GPT-5.5
53.5
66.4
Table 3: WeaveBench results on the full 114-task suite using GPT-5.5 at high reasoning effort and a 200-step agent limit. PassRate is the fraction of tasks with score at least 0.80 ; Overall is the mean per-task score.
IoU (%) ↑
Voxel IoU by difficulty (%) ↑
Mean
Easy
Medium
Hard
no tool
GPT-5.5, official score
44.4
–
–
–
GPT-5.5, ours reproduced
43.8 ( Δ−0.6 )
43.4
44.6
43.3
with tool
GPT-5.5, official score
55.8
–
–
–
Table 4: BenchCAD Vision2Code results across different difficulty levels; higher is better. ”No tool” mode tests the agents ability to output the CAD artifacts in a single agent turn. “With tool” mode provides the model a Python sandbox with the task images and CAD libraries, allowing it to render, measure, and iteratively refine its CAD program before submitting the final answer. The official leaderboard does not provide the breakdown scores for different difficulty levels, therefore we left it to be empty.
Overall (%) ↑
Generation (%) ↑
Editing (%) ↑
GPT-5.5
Codex + Build123d-MCP tools
45.7
37.2
58.8
CUAWright
49.4 ( Δ +3.7)
40.2 ( Δ +3.0)
63.5 ( Δ +4.7)
GPT-5.6
Codex + Build123d-MCP tools
53.2
45.1
65.6
CUAWright
55.7 ( Δ +2.5)
49.0 ( Δ +3.9)
66.0 ( Δ +0.4)
Table 5: CadGenBench results on all 81 public fixtures. We report the validated baseline score from the official benchmark using Codex with registered Build123d MCP tools. CUAWright uses the Build123d package through CLI calls in the terminal.
Execution mode
Tasks, n (%)
Mean partial (%) ↑
Exact completion (%) ↑
Terminal-only
30 (28.6)
71.89
33.3 (10/30)
Mixed terminal + GUI
75 (71.4)
66.32
25.3 (19/75)
Overall
105 (100.0)
67.91
27.6 (29/105)
Table 6: OSWorld 2.0 strict terminal-only versus mixed execution for the GPT-5.6 Sol Max Terminal-only traces contain no xdotool, pyautogui, or equivalent operating-system input invocation.
Figure 4: Top 5 failure reasons of GPT-5.5 and GPT-5.6 Sol run on OSWorld 2.0.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Metric
GPT-5.5 xhigh
GPT-5.6 Sol max
Tasks scored
105/105
105/105
Partial score
63.2242%
67.9130%
Exact passes
27/105
29/105
Input tokens
400,638,827
434,755,823
Output tokens
14,967,422
4,394,423
Total tokens
415,606,249
439,150,246
Appendix
Table 7: OSWorld 2.0 evaluation statistics and cost comparison.
Figure 5: Tool-crafted batched visual inspection in Task005. The agent writes and executes a lightweight sheet() utility in the terminal (left), combining 13 JPEG documents into one filename-annotated contact sheet (center). A single multimodal inspection replaces repeated GUI-based image opening and produces a structured response grouping the medical evidence by visit date, diagnosis, expenses, and supporting files (right).
Figure 6: Trace-backed root-cause analysis in WeaveBench. The agent first exports and analyzes the Jaeger traces programmatically, then drives the native UI to verify the causal path visually. The slow-trace waterfall (left) shows payments.gateway.http occupying almost the full 4.12-second checkout request. Expanding that leaf span (right) reads back the decisive evidence: HTTP 504, OpenTelemetry ERROR , an upstream timeout, and peer.service=payments-gateway .
Figure 7: Tool-crafted lookup commands reused across a task in Task008. The task (top) asks for a reimbursement claim whose evidence lives in a mail client and a bank account; the screenshot below it is the bank’s transactions tab after the agent’s own search tool has typed CATHAY . The agent writes mail_open_info.py and vault_search.py once (middle) and invokes them 13 times with the arguments shown (bottom), once per supporting e-mail and once per charge.
Figure 8: Tool-crafted click primitive reused across a task in Task070. The task (top) asks for an AWS billing anomaly to be found, fixed, and reported in TeamChat; the console’s cost panel below it already shows the spike. The agent writes cdp_click.py once (middle) and invokes it 15 more times with the arguments shown (bottom). The trajectory attaches no screenshot; the console view is taken from the GPT-5.6 run of the same task.
Figure 9: BenchCAD heat-sink reconstruction. Reference views appear on the left, the final render in the center, and an amplified pixel difference on the right. The submitted STEP file received voxel IoU 1.0000.
Task 045: multi-agent advantage
Task 076: direct advantage
Task
Record all seven dashboard pages in a 1920×1080 GIF lasting at least five seconds and smaller than one MiB
Apply the speaker-note instructions across six slides, including layout, sensor, chart, workbook, and table edits
Figure 10: Paired traces for tasks 045 and 076. In task 045, Critic feedback changed the artifact from seven selected frames to a 156-frame recording. In task 076, three inconclusive Verifier calls exhausted the shared budget and blocked submission.