PaperlayerView exact revision on arXiv

Source-linked AI summary

WorkBorne: Discovering Agentic Tasks from Real-World Artifact Transitions

Junyi Zhou, Yufei Wang, Qiyuan Zhang, Chuhan Wu, Zhengxi Li, Wentao Zhao, Jie Zhou, Le Tian, Xiao Zhou, Chen Ma

arXiv:2610.14915v1cs.AIcs.CL

TL;DR

Long-horizon agent tasks are scarce, while existing synthesis approaches struggle to combine complex instructions with authentic workspaces. WorkBorne discovers tasks through artifact transitions and instantiates them as grounded environments. WorkBorne-27B improves reported benchmark results over its base model.

  • Problem

    Long-horizon tasks are rare, and existing synthesis methods struggle to couple complex instructions with authentic, interactive environments.

    Source: §1 · pp. 1–2

  • Method

    WorkBorne constructs artifact corpora, discovers transitions linking personas, input artifacts, and output anchors, then instantiates validated transitions as tasks.

    Source: §1 · p. 2, §3.1 · p. 4, §3.2 · p. 4

  • Results

    WorkBorne-27B improves Terminal-Bench 2.1 accuracy from 43.1% to 61.8% and ClawEval General score from 75.0 to 80.2.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Transition modeling is strongly associated with execution depth, while corpus breadth and transition depth provide complementary scaling gains.

    Source: §1 · p. 2

Abstract

from arXiv · show

Developing terminal-based agents for long-horizon execution demands synthetic tasks that tightly couple complex instructions with authentic, interactive environments. However, current automated synthesis suffers from a fundamental asymmetry: instruction-oriented methods fabricate low-fidelity, synthetic environments, whereas resource-oriented methods treat real artifacts merely as static inputs, degrading task instructions into shallow, trivial operations. To address this limitation, we introduce WorkBorne, a synthesis framework that uses artifact transitions as the central primitive for modeling the full life-cycle of agentic tasks. In WorkBorne, each transition specifies who performs the task, which artifacts anchor the initial state and intended outcome, and the corresponding relationship between them. Specifically, WorkBorne first constructs coherent and diverse artifact corpora from iterative web retrieval, then discovers non-trivial transitions by identifying meaningful input-output artifact assemblies and the relations that connect them. Each discovered transition is finally instantiated into grounded agentic tasks. We construct 1,061 artifact corpora and 7,882 task environments, yielding 15,485 high-quality trajectories, with 53.6% exceeding 30 assistant turns. Experiments on Qwen3.5-27B and Qwen3.5-35B-A3B show that WorkBorne substantially improves over base models (i.e., improving Terminal-Bench 2.1 from 43.1% to 61.8% and ClawEval General score from 75.0 to 80.2) and achieves the strongest overall performance among other synthesis baselines (e.g., TMax (Ivison et al., 2026), RST (Li et al., 2026a), Terminal-Universe (Wu et al., 2026a)). Further analysis shows that transition is strongly associated with execution depth, while corpus breadth and transition depth provide complementary scaling gains.

1 Introduction

Long-horizon agent tasks are scarce, and existing synthesis methods struggle to combine authentic workspaces with complex instructions. WorkBorne models tasks as transitions between real artifacts and reports gains in task depth and benchmark performance.

1 Introduction

  • Long-horizon task instructions and environments must jointly induce sustained interaction, but tasks that genuinely do so remain exceptionally rare.Ordinary tasks tend to be shallow, and human instructions often minimize interaction steps.Source: §1 · p. 1
  • Instruction-oriented synthesis can yield low-fidelity environments, while resource-oriented methods use authentic artifacts as passive inputs and often produce shallow tasks.Together, these approaches struggle to combine authentic workspace richness with instruction complexity.Source: §1 · pp. 1–2
  • WorkBorne treats a task as a transition connecting a persona, real input artifacts, intended output artifacts, and their relationship.It constructs artifact corpora, discovers transitions, then instantiates them as grounded tasks.Source: §1 · p. 2
  • 1,061 artifact corpora and 7,882 executable tasks yielded 15,485 trajectories; WorkBorne-27B improved Terminal-Bench 2.1 accuracy from 43.1% to 61.8% and ClawEval General score from 75.0 to 80.2.53.6% of trajectories exceeded 30 turns; transition modeling also increased average execution length and was associated with execution depth and complementary scaling gains from corpus breadth and transition depth.Source: §1 · p. 2

2 Related Work

Terminal-agent research has shifted toward sustained, long-horizon execution, making scalable task synthesis a central challenge. Existing synthesis approaches trade off instruction complexity against authentic workspace structure.

2.1 Terminal-Based Agents.

  • As terminal agents move toward open-ended, long-horizon work, scalable data synthesis must preserve structural complexity and execution fidelity.The cited progression runs from interactive code execution to repository-scale engineering and general-purpose command-line interaction.Source: §2.1 · p. 2

2.2 Terminal Task Synthesis.

  • Instruction-oriented methods build environments around task specifications, while resource-oriented methods start from real workspaces but often yield shallow inspection tasks.Hybrid approaches improve grounding but still treat instructions and environments as decoupled synthesis targets.Source: §2.2 · pp. 2–3
  • WorkBorne instead organizes synthesis as artifact corpus construction, transition discovery, and task instantiation.The stages build real-world artifact pools, discover grounded transitions through assembly, planning, and validation, then compile aligned task components.Source: Figure 1 · p. 3

3 WorkBorne

WorkBorne formalizes terminal tasks as transitions over real artifacts, then discovers and instantiates validated transitions into executable tasks. Its collected data includes thousands of tasks and sustained trajectories.

3.1 Task Formulation

  • An artifact transition formalizes who performs the task, which input artifacts define the initial workspace, which output anchors specify the intended outcome, and how they relate.WorkBorne first discovers valid transitions from raw artifact spaces, then instantiates them as aligned instructions, environments, and rubrics.Source: §3.1 · pp. 3–4

3.2 Transition-Driven Task Synthesis

  • Corpora are built from skill-guided web retrieval, iteratively expanded to fill missing file formats, and indexed as file cards for transition discovery.The collection contains 1,061 curated skills spanning professional scenarios and multi-format artifacts.Source: §3.2 · p. 4
  • Transition discovery assembles persona-relevant inputs and real-world output anchors, plans how to adapt or integrate the inputs, and validates that each task requires multi-step execution.Invalid transitions are refined or discarded; validated transitions are jointly instantiated as instructions, environments, and evaluation rubrics grounded in the same transition.Source: §3.2 · pp. 4–5

3.3 Data Collection

  • Trajectory collection used teacher rollouts across 14 agent harnesses, filtered infrastructure failures, and retained trajectories scoring at least 8 out of 10 on task rubrics.Each task environment averaged 350 files and 10.4 rubrics, and 98.8% of file occurrences came directly from source corpora.Source: §3.3 · pp. 5–6
  • Across 1,061 corpora, retained trajectories averaged 37.5 assistant turns, with a median of 32; 53.6% exceeded 30 turns and 21.0% exceeded 50 turns, reflecting sustained, long-horizon interactions.Source: §3.3 · p. 6

4 Experiments

Experiments evaluate WorkBorne-trained models on Terminal-Bench Lite, Terminal-Bench 2.1, and ClawEval, including comparisons with synthesis methods and tests of generalization and model capacity. WorkBorne-27B improves over its base model across benchmarks and compares favorably with larger open models on Terminal-Bench 2.1.

4.1 Experimental Setup

  • Models are fine-tuned on all 15,485 WorkBorne trajectories for three epochs, and evaluated on Terminal-Bench Lite, Terminal-Bench 2.1, and ClawEval General.Terminal-Bench scores use three-run averages; the evaluation uses Terminus2 for both Terminal-Bench benchmarks and the official harness for ClawEval.Source: §4.1 · p. 6, §4 · p. 6

4.2 Main Results

  • 61.8% on Terminal-Bench 2.1, an 18.7-point gain over Qwen3.5-27B, matches DeepSeek-V4-flash-Preview and exceeds Kimi-K2.6 at 59.9%.WorkBorne-27B also reaches 84.3% on Terminal-Bench Lite, improving 16.0 points over its base model.Source: §4.2 · p. 6, Table 2 · p. 7
  • +4.4% pass@3, +5.2% overall score, and +3.1% strict joint accuracy (pass∧3) on ClawEval show gains across heterogeneous agent workflows.Terminal-Bench evaluates terminal execution via Terminus2, while ClawEval tests broader agent capabilities under claw-agent.Source: §4.2 · p. 6
  • Qwen3.5-27B gains +18.7% on Terminal-Bench 2.1, versus +8.6% for Qwen3.5-35B-A3B, from the same 15K-trajectory budget.The paper associates the larger dense-model gains on complex, long-horizon tasks with greater active capacity.Source: §4.2 · pp. 6–7

5 Analysis and Ablation Studies

The analyses test how transition modeling affects execution and training, which task properties relate to execution depth, and how corpus breadth and transition depth scale supervision. Transition modeling produces deeper execution and stronger downstream performance; scaling results favor broader corpus coverage at matched budgets while showing gains from additional transitions.

5. Analysis and Ablation Studies

  • The ablation compares Instruction-first, Input-only, and WorkBorne pipelines from a shared pool of 125 skills, using 1,000 environments and trajectories per pipeline.The analyses use matched trajectory budgets; the scaling study varies corpora and transitions per corpus across 1K, 3K, 5K, and 8K trajectories.Source: §5.1 · pp. 7–8, §5 · p. 7, §5.3 · p. 9

5.1 Construction Pipeline Ablation

  • 66.6% of WorkBorne trajectories exceed 30 turns, versus 33.0% for Input-only and 27.4% for Instruction-first; WorkBorne-27B reaches 71.1% on TB Lite and 51.7% on TB 2.1.With 1,000 training trajectories, those scores exceed Input-only by +2.1 and +3.4 points, and Instruction-first by +2.8 and +4.5 points.Source: §5.1 · p. 8, Table 3 · p. 8

5.2 Transition Structure and Execution Depth

  • Artifact assembly associates more strongly with execution depth (η = 0.314) than input or total artifact counts (η = 0.049 and 0.015), while output form exceeds input form (0.306 vs. 0.183).Within a fixed repository-to-repository assembly, median execution is 47 turns for porting and 34 for structural imitation.Source: §5.2 · p. 8, Figure 3 · p. 9

5.3 Corpus and Transition Scaling

  • Both scaling paths improve performance; increasing transitions per corpus from 1 to 8 raises TB Lite from 72.5% to 79.4% and TB 2.1 from 52.8% to 59.6%.At every matched budget before the shared endpoint, spreading trajectories across more corpora performs better than sampling more transitions from fewer corpora.Source: §5.3 · pp. 9–10, Figure 4 · p. 9

6 Conclusion

The conclusion presents WorkBorne as a framework for synthesizing terminal-agent tasks through artifact transitions, which connect authentic workspaces with substantive task instructions. Its analyses associate execution depth more with transition structure than artifact quantity.

6 Conclusion

  • WorkBorne models agentic tasks through artifact transitions, using a 3W primitive to specify who works, which inputs are provided, and which output anchors are intended.The conclusion reports improved performance across benchmarks and says execution depth is associated primarily with transition structure rather than artifact quantity.Source: §6 · p. 10

AI Use Statement

The authors report using generative AI for language polishing and as part of the WorkBorne pipeline to generate tasks and trajectories, with author review and validation. The retained training data spans 14 harnesses without one interface dominating.

AI Use Statement

  • Generative AI assisted with manuscript polishing and WorkBorne task and trajectory generation; authors reviewed and validated the AI-assisted content and data.The authors state they take full responsibility for the final content.Source: AI Use Statement · p. 10

A Harness Statistics

  • Each of the 14 harnesses contributes at least 555 retained trajectories, and the largest share is 10.4%.The statistics cover 15,485 training records after rubric filtering and cleaning.Source: A Harness Statistics · p. 15

B Data Contamination Checks

WorkBorne checks for benchmark contamination at the task, trajectory, and artifact levels, and reports no detected overlap with the evaluated benchmarks.

B Data Contamination Checks

  • No task overlap or benchmark leakage was found, and maximum instruction-level Jaccard similarity was below 0.013 across all three benchmarks.The checks compare training instructions, metadata, and trajectories with evaluation instructions, then inspect artifact selections and environment records.Source: B Data Contamination Checks · pp. 15–16, Table 6 · p. 16
  • Artifact contents and provenance metadata were screened for benchmark identifiers and overlap signals, with artifacts showing leakage excluded.Retrieved artifacts also underwent license-aware filtering and handling based on available license information.Source: B Data Contamination Checks · p. 16

C Training Configurations

Both base models use the same WorkBorne training data and shared optimization settings, with model-specific MoE handling for Qwen3.5-35B-A3B.

C Training Configurations

  • Qwen3.5-27B and Qwen3.5-35B-A3B use the same WorkBorne data and optimization settings for comparable training-data effects.Both receive full-parameter SFT with MS-SWIFT 4.5.2 and Megatron/MCore, while the vision encoder and aligner remain frozen.Source: C Training Configurations · p. 16, Table 7 · p. 17
  • Qwen3.5-35B-A3B additionally uses expert parallelism and an auxiliary load-balancing loss.Its MoE settings are listed separately from the shared settings.Source: C Training Configurations · p. 17, Table 8 · p. 17
  • Trajectories longer than 256K tokens are discarded without truncation; retained examples preserve reasoning, supervise assistant responses, and separate attention across packed trajectories.The MoE configuration also enables all-to-all token dispatch, grouped GEMM, permutation fusion, and shared-expert overlap, with dropless expert capacity.Source: C Training Configurations · p. 17

D Evaluation Configurations

The evaluation protocols specify model-only assessment settings for Terminal-Bench and ClawEval; construction statistics define environment and tool-call measures and their coverage.

D Evaluation Configurations

  • Terminal-Bench Lite reports mean raw reward and Terminal-Bench 2.1 reports pass@1, each averaged over three runs using Terminus2.Evaluation uses interleaved thinking, top-k = 20, context compression at 16,384 remaining tokens, four-hour task timeouts, and official task resources; partial Lite rewards are not binarized.Source: D Evaluation Configurations · p. 17, Table 9 · p. 18
  • ClawEval assesses its General subset with the endpoint’s default thinking setting and Gemini 3 Flash Preview via OpenRouter as the judge.Both evaluation pipelines use the language-model-only option.Source: D Evaluation Configurations · pp. 17–18

E Construction Comparison Statistics

  • Construction-comparison environment metrics summarize 1,000 environments per pipeline, while execution metrics summarize retained trajectories whose task coverage may differ.Environment size is measured in logical bytes; file types count suffixes, maximum path depth is averaged across environments, and file categories count distinct recorded categories with files.Source: E Construction Comparison Statistics · p. 18, Table 10 · p. 18
  • Tool-call means use fully parsed action records from 993 Instruction-first, 997 Input-only, and 996 WorkBorne trajectories, excluding malformed records only for this metric.Counts treat each structured invocation once; Terminus2 command or keystroke entries, including polling, count individually, while shell invocations are not split into constituent commands.Source: E Construction Comparison Statistics · p. 18

F Human Validation of Task and Rubric Quality

A blinded matched-task review rates WorkBorne higher in task complexity, workspace realism, and instruction-environment alignment, while its synthesized rubrics also receive high ratings.

F Human Validation of Task and Rubric Quality

  • Two expert reviewers independently rated 150 blinded tasks in 50 matched triplets on complexity, workspace realism and coherence, and instruction-environment alignment.The 50 sampled WorkBorne tasks also received rubric ratings for requirement coverage, faithfulness and correctness, and verifiability; scores were averaged across reviewers.Source: F Human Validation of Task and Rubric… · pp. 18–19
  • 4.21 versus 2.94 and 2.61 for Task Complexity, WorkBorne also scores 4.46 in Workspace Realism & Coherence and 4.35 in Instruction-Environment Alignment.These human ratings corroborate execution-based trends and indicate that greater execution demands do not come from incoherent workspaces or poorly grounded instructions.Source: Table 11 · p. 19, F Human Validation of Task and Rubric… · p. 19
  • 84.0–92.0% of ratings are at least 4 across the three rubric dimensions, with WorkBorne rubrics averaging 4.35 overall.The audit covers the rubric sets for 50 sampled WorkBorne tasks.Source: Table 12 · p. 19, F Human Validation of Task and Rubric… · p. 19

G Task and Artifact Distributions

The distributions show that corpora support multiple task directions and that workspaces preserve heterogeneous source materials. Execution depth varies with transition structure, while artifact counts alone have near-zero correlation with it.

G Task and Artifact Distributions

  • A corpus yields a median of eight tasks, and 86.0% contain at least five transition types; five frequent types account for 77.2% of tasks.The five are implementation from a specification, requirement evolution, porting, structural imitation, and reconciliation.Source: G Task and Artifact Distributions · p. 19, Figure 5 · p. 20
  • 98.8% of file occurrences are copied from source corpora; the median workspace contains 98 files across 15 extensions, and source code makes up 39.6% of files.Documents occur in 98.0% of tasks and structured data in 82.4%; the 90th-percentile workspace contains 1,034 files.Source: G Task and Artifact Distributions · p. 19, Figure 5 · p. 20
  • Directory trees appear among output anchors in 67.1% of tasks versus 33.2% of inputs, while documents appear in 47.9% of inputs versus 23.7% of output anchors.The authors interpret the directory-tree pattern as consistent with output anchors supplying workspace structure to guide the intended outcome.Source: G Task and Artifact Distributions · pp. 19–20

H Transition Structure and Execution Comparisons

  • Among tasks with at least two cases, 34.8% of 396 repository/repository tasks exceed 50 median turns, versus 11.7% of 383 document/document tasks.These are descriptive comparisons on retained data, and task and corpus composition varies across groups.Source: H Transition Structure and Execution… · p. 21, Table 13 · p. 21
  • Execution depth has near-zero Spearman correlation with input count (ρ = 0.010) and total selected artifact count (ρ = 0.004).The authors report that artifact assemblies and planned relations provide more informative distinctions in execution depth than artifact counts alone.Source: H Transition Structure and Execution… · p. 21
  • A 73-turn port of CBraid from C++ to Python combined implementation, repeated comparison with the C++ outputs, regression testing, and verification, ending with 280 passing tests.The trajectory used pyknotid and braid-visualiser as output anchors for package structure and class interfaces.Source: H Transition Structure and Execution… · p. 21

Retrieval procedure

  • Retrieval seeks verified, complementary artifacts for a workflow, then assembles disjoint inputs and output anchors for personas whose work fits the corpus.Retrieval favors stable sources, records provenance and known licensing, and excludes known evaluation materials or resources whose terms prohibit the intended use.Source: Retrieval procedure · p. 22, §1 · p. 22, §2 · p. 23

4. Task Instantiation

  • Task instantiation aligns the user instruction, initial environment, and evaluation criteria with the approved transition and its requirements.The templates combine instruction generation, environment construction, and rubric synthesis in one invocation; criteria must follow explicit requirements or necessary conditions for a correct, usable result and preserve valid alternatives.Source: I Prompt Templates · pp. 21–22, §4 · p. 25, Evaluation criteria · p. 26

Related papers

Loading 2610.14915v1…