PaperlayerView exact revision on arXiv

Source-linked AI summary

CUA-Universe: A Scalable and Dynamic Environment for Hybrid GUI+CLI Agents

Haoting Shi, Wenhao Wang, Weicheng Fang, Yaozhong Liang, Tian Jin, Pengxiang Zhao, Guangyi Liu, Siheng Chen, Yanfeng Wang

arXiv:2609.05374v1cs.AI

TL;DR

Scalable hybrid GUI+CLI environments for real software remain scarce, and agents still need to coordinate visual interaction with precise command-line operations. CUA-Universe creates a pipeline for building such environments and hybrid training data; its 9B model improves success and efficiency across CUA-Verse, OSWorld, and OSWorld-MCP.

  • Problem

    Scalable real-software GUI+CLI environments are scarce, while agents still predominantly use GUI and need to coordinate visual inspection with precise CLI operations.

    Source: §1 · p. 1

  • Method

    CUA-Universe builds a scalable pipeline that adapts real desktop apps into reproducible GUI+CLI environments and synthesizes controllable hybrid tasks and efficient-path trajectories.

    Source: §1 · p. 1, §3.1 · p. 3

  • Results

    Across CUA-Verse, OSWorld, and OSWorld-MCP, the 9B model improves success and efficiency; on CUA-Verse it reaches roughly 3× base-model success with 60% fewer tokens.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The results indicate that hybrid interaction skills can be learned and transfer beyond CUA-Verse to OSWorld and an unseen tool interface.

    Source: §5 · p. 7

    The pipeline currently covers single-application tasks; cross-application workflows, with state carried across apps, remain future work.

    Source: K Limitations · p. 18

Abstract

from arXiv · show

Computer-use agents have advanced on benchmarks like OSWorld and AndroidWorld, but still act mostly through the GUI, often producing inefficient trajectories. Real-world computer work is hybrid, combining visual-state inspection with precise, high-throughput command-line operations, so capable agents must coordinate both modalities over shared application state. Yet scalable hybrid environments remain scarce because supporting both GUI and CLI over real applications typically requires substantial manual engineering for each application. Existing agents also struggle to use the two interfaces complementarily: CLI-native agents lack visual perception for tasks involving interface state or layout, while GUI-native agents are inefficient for operations better executed through commands. We introduce CUA-Universe, a scalable environment-to-data pipeline that turns real desktop software into hybrid GUI+CLI environments. App-Forge adapts applications into reproducible VMs and command-line surfaces it discovers, wraps, or generates, scaling to 16 applications; Task-Weave synthesizes diverse hybrid tasks of controllable difficulty from reusable operations over seed files; and Path-Steer steers rollouts along efficient hybrid paths and harvests verified trajectories for post-training. Training on this data shifts behavior from inefficient GUI interaction and brittle CLI scripting toward effective GUI+CLI orchestration. Our 9B model improves both success and efficiency on CUA-Verse (Score +39.3 pts; -37% steps, -60% tokens), OSWorld (SR +16.8 pts; -57% steps, -44% tokens), and OSWorld-MCP (Score +7.84 pts; -27% steps, -30% tokens). CUA-Universe provides a scalable path toward more capable and efficient computer-use agents.

1 Introduction

CUA-Universe converts real desktop software into scalable hybrid GUI+CLI environments and training data, targeting agents that must coordinate both interfaces over shared application state. Its 9B model shows stronger and more efficient performance across CUA-Verse, OSWorld, and OSWorld-MCP.

1 Introduction

  • GUI-only trajectories can be unnecessarily long, while single-modality agents struggle with complementary work: CLI agents lack visual perception and GUI agents waste effort on command-suited batch operations.The central challenge is choosing when to switch interfaces and carrying task state across them, such as locating a target through the GUI before processing it through the CLI.Source: §1 · p. 1
  • App-Forge, Task-Weave, and Path-Steer adapt applications, synthesize controllable hybrid tasks, and harvest efficient trajectories, scaling the pipeline to 16 desktop applications.App-Forge discovers, wraps, or generates application-specific command-line surfaces for reproducible VMs; Task-Weave builds tasks from reusable operations and seed states; Path-Steer steers rollouts toward efficient hybrid paths.Source: §1 · pp. 1–2
  • CUA-Verse Score improved +39.3 pts with −37% steps and −60% tokens; OSWorld SR rose +16.8 pts with −57% steps and −44% tokens; OSWorld-MCP Score gained +7.84 pts with −27% steps and −30% tokens.The introduction also reports roughly 3× the base model’s success on CUA-Verse, a +16.8-point OSWorld gain over GUI-only execution, and a +7.84-point OSWorld-MCP gain over the base model.Source: Figure 1 · p. 2, §1 · p. 2
  • CUA-Verse evaluates orchestration over shared application state with 160 held-out hybrid tasks across eight desktop applications introduced by CUA-Universe.The tasks are disjoint from training data, and the model’s learned hybrid skills also transfer to OSWorld and the unseen tool-invocation interface of OSWorld-MCP.Source: §1 · p. 2

2 Related Work

Prior environment-synthesis and GUI/CLI-agent work advances automation within single modalities or fixed hybrid setups; CUA-Universe grounds hybrid tasks and tools in real desktop applications and steers efficient trajectories.

2.1 Environment Synthesis for Computer-Use Agents

  • Prior synthesized environments range from artificial GUI-only web pages to desktop and mock-web applications, but remain GUI-driven; CUA-Universe creates shared-state GUI+CLI tasks with orchestration signals.Its hybrid trajectories provide efficiency and orchestration signals that GUI-only pipelines cannot structurally provide.Source: §2.1 · p. 3

2.2 Single-Modality GUI and CLI Agents

  • Terminal agents miss visual-state operations, while GUI agents handle batch operations slowly; hybrid training teaches modality switching over shared application state.The passage describes GUI agents as also inefficient on long-horizon, cross-application workflows.Source: §2.2 · p. 3

2.3 Toward Hybrid GUI+CLI Agents

  • Unlike hybrid agents built around fixed tools and task sets, CUA-Universe grounds tools in real desktop applications, generates agent-native CLIs when needed, and steers efficient paths for post-training.Path-Steer steers rollouts onto efficient hybrid paths, raising success rates and yielding higher-quality trajectories.Source: §2.3 · p. 3

3 Method

CUA-Universe builds reproducible hybrid GUI+CLI environments, synthesizes seed-grounded tasks, and collects efficient verified trajectories. It models tasks as partially observable interactions over a shared persistent application state, with a verifier scoring each trajectory.

3.1 Overview and Problem Formulation

  • The task formulation treats each hybrid task as a partially observable process in which GUI and CLI actions read and modify shared persistent application state.The task tuple comprises an instruction, seed initial state, and verifier that scores the resulting trajectory.Source: §3.1 · p. 3

3.2 App-Forge: Scalable Agentic Environment Construction

  • App-Forge adapts desktop applications into reproducible environments and supplies CLI surfaces aligned with their GUIs, scaling the pipeline to 16 applications.It reuses an installation workflow and combines discovered native tools, wrapped scripting APIs, and generated CLIs; adapters resynchronize GUI views after CLI edits to shared project state.Source: §3.2 · pp. 3–4

3.3 Task-Weave: Compositional Hybrid Task Synthesis

  • Task-Weave builds reusable operations from GUI exploration, composes them into seed-conditioned tasks of controlled difficulty, and validates them through real execution.Tasks package a natural-language goal, initial state, verifier, and guidance; a review agent checks feasibility, ambiguity, and whether the seed already satisfies the goal, revising fixable tasks and discarding invalid ones.Source: §3.3 · p. 4

3.4 Path-Steer: Efficiency-Aware Hybrid Rollout

  • Path-Steer supports interleaved GUI and CLI actions, guides modality choices by operation type, and retains high-scoring verified rollouts for post-training.Its execution prior favors CLI for batch, precise, or high-throughput operations and GUI for tasks requiring visual layout or interface state. The pipeline produces 4,923 episodes and approximately 235K training steps across 16 applications.Source: §3.4 · p. 4, Figure 2 · p. 5

4 Experiments

Experiments assess CUA-Universe training on CUA-Verse and its transfer to OSWorld and OSWorld-MCP, while separately testing Path-Steer during data generation. The results show improvements in task performance and efficiency across these evaluations.

4 Experiments

  • 4,923 verified episodes, totaling about 235K step-level records, train Qwen3.5-9B through three epochs of LoRA fine-tuning.The data comes from Kimi K2.5 rollouts under Path-Steer, retaining trajectories that pass a VLM judge at a 0.75 score threshold.Source: §4 · p. 5

4.1 Evaluation on CUA-Verse

  • 0.582 Score makes Ours the top open-source CUA-Verse model, above Kimi K2.5 (0.522) and behind Seed2.1 Pro (0.599) and GPT-5.5 (0.768).Against identical Qwen3.5-9B (0.189), Ours improves Score about 3× with 37% fewer steps and 60% fewer tokens (255K vs. 643K per episode). It is strongest on Audacity (0.815) and OBS (0.605), and weaker on Blender (0.398) and Godot (0.460).Source: §4.1 · p. 5, Figure 3 · p. 5, Table 1 · p. 6

4.2 Transfer to OSWorld

  • +16.8 points raises Ours’ OSWorld SR from 23.4% with GUI-only execution to 40.2% with GUI+CLI, within 3 points of EvoCUA-8B (43.3%).Ours uses the fewest tokens per task (286.5k) and steps (28.6) of any agent; adding CLI converts +41 previously failed GUI tasks into successes, and Ours uses 2.35× fewer steps on jointly solved tasks. The comparison uses the same 244 tasks, task text, environment, and verifier across interfaces.Source: §4.2 · pp. 5–6, Figure 4 · p. 6, Table 1 · p. 6

4.3 Generalization to OSWorld-MCP

  • +7.79 points improves Ours’ OSWorld-MCP SR from 20.90% to 28.69% over the base, while TIR rises from 10.66% to 23.36%.Despite training only with application-specific CLI tools and never seeing the MCP action space, Ours scores 29.51% versus Seed2.1 Pro’s 27.03%; it reaches ACS 27.25 with 87.95M tokens, reducing steps by 27% and tokens by 30% relative to its base.Source: Table 2 · p. 6, §4.2 · p. 6, §4.3 · p. 7

4.4 Rollout Efficiency

  • 0.44 to 0.51 Accept Rate shows Path-Steer improves Kimi K2.5 rollout quality while reducing steps from 26.68 to 22.75.Mean Score rises from 0.63 to 0.71, tokens fall from 385K to 332K, and per-task cost falls from $0.31 to $0.26. Seed2.1 Pro shows the same pattern: Accept Rate improves from 0.45 to 0.54 and Mean Score from 0.67 to 0.75, with lower steps, tokens, and cost; both settings have identical CLI access, isolating the contribution of steering.Source: §4.4 · p. 7, Table 3 · p. 7

5 Conclusion · A Additional Experiments · A.1 Action-Modality Behavior on CUA-Verse

CUA-Universe scales shared-state GUI+CLI environments and trajectories, and its 9B model demonstrates learnable, transferable hybrid interaction. On CUA-Verse, results indicate that effective performance depends on coordinating modalities, not simply maximizing CLI use.

5 Conclusion

  • CUA-Universe scales computer-use learning by turning real desktop software into shared-state GUI+CLI environments that provide hybrid tasks and trajectories.The paper frames this as scaling environments and interaction spaces rather than relying only on larger models or more GUI-only trajectories.Source: §5 · p. 7
  • The resulting 9B model shows transferable hybrid capability, with strong gains on CUA-Verse and OSWorld and further generalization to the unseen OSWorld-MCP tool interface.The conclusion presents hybrid environment construction as a scalable source of learning across applications, tools, tasks, and interaction strategies.Source: §5 · p. 7
  • The paper proposes shifting CUA development from collecting ever more single-modality behavior toward scalable environments that continuously generate richer supervision.The stated aim is to support more capable, efficient, and general computer-use agents.Source: §5 · p. 7

A.1 Action-Modality Behavior on CUA-Verse

  • CUA-Verse tasks require both GUI and CLI operations over shared application state, including visually grounded interactions that cannot be reduced to command execution alone.The analysis classifies each nonterminal action as GUI or CLI and reports CLI share alongside Steps and Score.Source: §A.1 · p. 10
  • Hybrid training raises CLI use to 25.3%, cuts mean Steps by 37%, and increases Score from 0.189 to 0.582 over the GUI-bound Qwen3.5-9B base.The same-backbone comparison indicates that the model learns to use CLI operations while retaining GUI interaction when needed.Source: §A.1 · p. 10, Table 4 · p. 10
  • GPT-5.5 uses CLI for all recorded actions and scores 0.768, yet falls short of perfect performance; GUI-based EvoCUA-8B scores only 0.330.The results indicate that CLI use alone is insufficient and that task-appropriate modality selection and coordination matter.Source: §A.1 · p. 10, Table 4 · p. 10
  • Table 5 reports OSWorld transfer success rates under a controlled 244-task GUI+CLI scope, comparing an 8-app OOD-only LoRA with a full 16-app LoRA.The OOD-only training applications are disjoint from OSWorld applications; final-row parenthesized values denote absolute percentage-point gains over untuned Qwen3.5-9B.Source: Table 5 · p. 10

A.2 Out-of-Domain Training Transfer

The 8-app OOD-only LoRA is tested zero-shot on a controlled 244-task GUI+CLI scope using applications disjoint from OSWorld’s eight evaluation applications. It improves overall success over the untuned base, while the full 16-app LoRA yields a larger gain.

A.2 Out-of-Domain Training Transfer

  • The 8-app OOD-only LoRA raises overall success rate from 24.2% to 27.4% (+3.2 points), while the full 16-app LoRA reaches 40.2% (+13.2 points) over the base.The OOD-only LoRA improves over the base on seven of eight applications; the full 16-app LoRA improves on all eight.Source: §A.2 · p. 10
  • OOD transfer varies by application: gains are largest on Thunderbird (+13.4 points) and VS Code (+8.7), while Impress decreases by 1.9 points.Writer improves by +4.3 points and GIMP by +3.9 points.Source: §A.2 · p. 10
  • The zero-shot gains indicate partially application-agnostic hybrid GUI-and-CLI coordination, while in-domain application coverage compounds that transferable capability.The paper reports that transferable hybrid-interaction skills and application-specific training data provide complementary benefits.Source: §A.2 · p. 10

B Adapted Applications … F.1 Agreement with Human Labels

This section describes how CUA-Universe adapts desktop applications with bounded, state-tied command surfaces and uses application-specific interfaces for seed-grounded task synthesis and guided GUI+CLI rollouts. It also reports human-label validation of the VLM judge used to score task success and filter training data.

B Adapted Applications

  • Eight inherited and eight added applications implement a uniform adapter interface exposing graphical, command-line, and programmatic access to the same underlying state.The supported application set spans diverse software domains.Source: B Adapted Applications · pp. 10–11, Table 6 · p. 11

C Tool Construction

  • Application adapters use distinct state paths: GIMP drives its live GUI; Blender runs bounded bpy against the real .blend; QGIS uses bounded PyQGIS with optional GUI sync; Zotero accesses its local database, browser connector, and Local API.Blender operations can inspect and permanently alter the native artifact; QGIS operations cover layers, features, layouts, and processing pipelines; Zotero supports management of collections, items, notes, and saved searches.Source: §C.1 · p. 11, §C.2 · p. 11, §C.3 · p. 12
  • ∼404 agent-visible commands span 16 applications—151 across eight OSWorld-reused apps and 253 across eight CUA-Universe extensions—through bounded JSON subcommands.A metadata registry stores command templates, while the rollout runner executes commands in the benchmark VM and returns exit code, stdout, and stderr; harness effects remain auditable in files opened by the GUI.Source: Figure 5 · p. 11, C Tool Construction · p. 11
  • OBS serializes a typed production model to native scene collections; Writer and Impress use ODF/UNO; Kdenlive edits native MLT XML projects.These surfaces support live production setup, rich-text editing, slide editing, and video timelines with clips, filters, transitions, and guides, respectively.Source: §C.5 · p. 12, §C.6 · p. 12, §C.12 · p. 13
  • Godot uses engine conventions and CLI runtime; draw.io exposes its native XML model through a constrained API; Audacity controls the running GUI project through Mod-Script-Pipe.Chrome provides a restricted DOM surface for page, accessibility-tree, and action commands, while VS Code uses its installed binary for file, extension, settings, and keybinding operations.Source: §C.8 · p. 13, §C.9 · p. 13, §C.10 · p. 13

C.13 VS Code — 13 commands; C.14 VLC — 13 commands

  • VS Code exposes workspace and extension commands, while VLC commands cover media playback, snapshots, transcoding, and configuration.The VLC surface also supports opening local or network media and launching media at specified times or segments.Source: §C.13 · pp. 13–14, §C.14 · p. 14

C.15 LibreOffice Calc — 10 commands

  • The Calc CLI implements bounded spreadsheet creation and manipulation through ODF/UNO sheet and cell operations.It is the Calc slice of the shared cli-anything-libreoffice harness and includes export presets and rendering.Source: §C.15 · p. 14

C.16 Thunderbird — 10 commands

  • Thunderbird commands use the installed executable’s profile-aware launch and compose capabilities to open mail, contacts, and calendars and compose email.Compose commands include CC, BCC, attachments, and mailto handling.Source: §C.16 · p. 14

D.1 Seed Examples

  • Seed files provide real, pre-loaded application states and verifiable content for synthesizing executable, diverse tasks with objective success criteria.A coding agent finds and downloads public source files, then normalizes them with format-conversion tools; Figure 6 shows examples across six application types.Source: §D.1 · p. 14, Figure 6 · p. 15

E.1 Rollout Agent Prompt

  • The rollout interface exposes GUI and CLI tools together, while Path-Steer adds an efficiency-aware plan to guide hybrid interleaving during data generation.CLI calls return structured JSON observations, and the guidance field is empty at evaluation.Source: §E.1 · p. 14

E.2 Steer-Path Rollout Examples

  • In the VLC task, Path-Steer helped the rollout recover from ineffective CLI attempts and complete the task, while the GUI-only workflow became stuck and failed.The steered trajectory switched to ffmpeg to rotate the video, then exported it and opened the Audio Effects panel; it did not necessarily reduce local trial-and-error.Source: §E.2 · p. 14

F VLM Judge

  • The VLM judge scores trajectories from task instructions, trajectory renderings, and screenshots, treating CLI output and exported artifacts as authoritative for file-producing tasks.It returns a score from 0 to 1 with a short justification; the final screenshot is primary evidence otherwise. The judge is GPT-5.4 at temperature 0.1 and uses up to 15 screenshots.Source: F VLM Judge · pp. 14–15

F.1 Agreement with Human Labels

  • 97.0% reweighted agreement (κ = 0.94) and 99.0% precision on accepted trajectories validate the judge, with no accepted human failures.The accepted set contained 480 trajectories, giving a label-noise upper bound of ≤1.0%; the rejection set had a 5.2% false-negative rate. Under a lenient criterion that counts partial successes as positive, precision was 100%, and the 480 accepted trajectories were all at least partial successes.Source: §F.1 · p. 15

G Rollout Harness and Budget · H Training Details

The rollout harness runs tasks in isolated OSWorld VMs under a shared action budget, while CPU-parallel collection makes the data-generation pipeline feasible on one ordinary server. The supplied passages describe rollout setup and budget, but contain no Training Details subsection content.

G Rollout Harness and Budget

  • Each task runs in an isolated OSWorld VM with 4 vCPU, 4 GB guest RAM, a copy-on-write overlay disk, and a 3 s post-action settle.All agents share a 60-step budget; Qwen-family models use a 3-frame history with at most 4 images per step, while other agents use their baseline defaults.Source: G Rollout Harness and Budget · pp. 15–16
  • The VLC trajectory with Path-Steer combines CLI tools and GUI actions, while the unguided trajectory gets stuck during filter configuration and fails.This figure illustrates the contrast between a guided hybrid trajectory and an unguided one.Source: Figure 7 · p. 16
  • ≈1.3–1.5 calendar days on one 128-core host sufficed in practice to collect the training rollouts, with the pipeline shardable across additional machines.A single rollout averages about 5 minutes; reset overhead and occasional VM stalls make throughput slightly below the estimated just-over-one-day ceiling.Source: G Rollout Harness and Budget · p. 16
  • A 128-core CPU server can host 32 environments concurrently, reducing the estimated 50,000 VM-minutes of rollout collection to about 1,560 minutes.Each environment needs about 4 vCPU and no GPU, making collection CPU-bound and embarrassingly parallel.Source: G Rollout Harness and Budget · p. 16

H.1 Data … I.1 Evaluation Agent Prompt

The section specifies the training-data pools and fine-tuning setup, then describes evaluation modes that test GUI and CLI use without task hints. It also illustrates how combining interfaces can help complete a settings task that GUI-only interaction fails.

H.1 Data

  • Every training record is a VLM-filtered step-level Qwen XML example with three-step history and a VLM score of at least 0.75.Source: §H.1 · p. 16
  • The CUA-Verse and OSWorld pools are balanced at 17,511 and 11,915 records per app, respectively, across eight applications each.Figure 2 summarizes pool splits and per-app episode counts; Table 8 reports final split sizes.Source: §H.1 · p. 16
  • OSWorld records have exactly four images, whereas 22,590 CUA-Verse records do not have four images.The CUA-Verse pool is VLM-filtered but does not enforce a four-image count.Source: §H.1 · p. 16

H.2 Base Model and Fine-tuning

  • Fine-tuning uses Qwen3.5-9B in bfloat16 with LoRA; the base weights, vision tower, and multimodal aligner are frozen, and only language-model linear modules train.Table 10 lists the optimization and sequence settings.Source: §H.2 · p. 16

I.1 Evaluation Agent Prompt

  • Each backbone uses its own actor prompt and runs in CLI mode with the execute cli function or GUI mode with the standard screenshot-and-pyautogui protocol.Figure 12 presents the Qwen CLI-mode prompt as a representative example.Source: §I.1 · p. 16
  • In the Figure 8 VS Code task, GUI+CLI completed the settings change in 5 CLI calls and 12 GUI actions (reward 1), while GUI-only exhausted 60 actions without committing it (reward 0).Source: Figure 8 · p. 17
  • At evaluation, path hints default to none, so policies receive no task or solution guidance; Path-Steer priors are used only during data-generation rollouts.Source: §I.1 · p. 17

I.2 CUA-Verse Benchmark

CUA-Verse is a benchmark of 160 hybrid GUI+CLI tasks across eight applications covered by the pipeline. Tasks mix modalities in every application, with CLI operations comprising most abstract tools on average.

I.2 CUA-Verse Benchmark

  • Blender is GUI-dominant, while Zotero and Kdenlive are CLI-heavy, reflecting differences in their task operations.Blender tasks involve spatial 3D manipulation; Zotero and Kdenlive tasks involve structured library and timeline edits.Source: §I.2 · p. 17
  • The eight benchmark applications are in-domain by design, measuring hybrid competence on software covered by the pipeline.Tasks are synthesized by the same pipeline as the training data.Source: §I.2 · p. 17
  • Tasks use 5.7 abstract tools on average, 59% of them CLI, and every application requires both GUI and CLI operations.The benchmark contains 160 tasks across eight applications, with 20 tasks per application; Figure 9 shows the per-application composition.Source: Figure 9 · p. 17, §I.2 · p. 17

J OSWorld Example

A VS Code task illustrates how GUI+CLI orchestration can combine visual discovery with a precise CLI edit, succeeding where GUI-only interaction fails.

J OSWorld Example

  • GUI+CLI completed the VS Code setting task with 5 CLI calls and 12 GUI actions (reward 1), while GUI-only exhausted its 60-action budget without committing the configuration (reward 0).The agent used the GUI to locate the setting, then wrote ”debug.focusEditorOnBreak”: false to settings.json via CLI and verified persistence.Source: J OSWorld Example · pp. 17–18

K Limitations

CUA-Universe is limited to single-application tasks and supervised fine-tuning, while some evaluation label noise remains. Its adaptation also assumes open-source or scriptable software on desktop Linux.

K Limitations

  • CUA-Universe synthesizes single-application tasks; cross-application workflows remain future work, though they are a natural extension of the pipeline.Single-application tasks enable verifiable environments and efficient hybrid paths at scale.Source: K Limitations · p. 18
  • Training uses supervised fine-tuning only, with verifier-based rewards for reinforcement learning identified as a next step toward surpassing the teacher.The pipeline already produces per-task verifiers that could serve as rewards.Source: K Limitations · p. 18
  • VLM-judged task success may contain label noise despite 99.0% acceptance precision and Cohen’s κ = 0.94 in judge–human validation.A conservative threshold and CLI and exported-artifact evidence mitigate, but do not eliminate, noise.Source: K Limitations · p. 18
  • Application adaptation assumes open-source or sufficiently scriptable software and desktop Linux; extending to closed-source software or non-desktop platforms remains open.The coding-agent-driven construction is not inherently tied to these choices.Source: K Limitations · p. 18

Related papers

Loading 2609.05374v1…