We study environment scaling as a way to increase agent-benchmark difficulty while holding core task objectives fixed. Our framework, Recursive Environment Scaling (RES), transforms existing benchmark environments while preserving their objectives and completion requirements. A proposer changes information access, resource relations, and execution dependencies; execution-grounded checks validate the resulting candidate, and accepted changes are inherited across rounds. This produces related environments for paired evaluation and training-trajectory collection. Experiments on 294 initial tasks from three benchmarks show lower task success for all six evaluation models, with decreases of 57.14--77.14 percentage points on Claw-Eval-Live. With equal trajectory counts and identical optimization settings, fine-tuning on RES trajectories outperforms original-trajectory SFT in all nine model--benchmark comparisons by 7.00--14.61 percentage points. Ablations identify recursive inheritance and dependency composition as important contributors to increased difficulty. Together, these results support environment scaling for both agent evaluation and training.
Figures & tables
Figure 1: Task scaling and recursive environment scaling. Task scaling expands tasks and trajectories. RES preserves core objectives while recursively transforming information access, resource relations, and execution dependencies. Accepted changes accumulate across rounds.
Figure 2: Environment scaling increases execution demand. Historical means for 105 Claw-Eval-Live tasks using DeepSeek-V4.1-Flash (one attempt, 50 turns, success threshold 0.80). Budget d averages retained states 0,…,d , carrying forward early-stopped finals (Appendix C ). Tokens include target input and output; reference nodes count added structure, not shortest solution paths.
Figure 3: Overview of RES. The proposer transforms accepted parent Ek . Execution-grounded checks validate the changes, task validity, and dependency feasibility, guiding revisions before accepting a child. Accepted children become parents; target evaluation separately informs stopping.
Claw-Eval-Live N=105
Workspace-Bench-Lite N=100
Terminal- Bench 2.1 N=89
Model
Environment
Pass@1
Score
TCR@80
RPR
Pass@1
DeepSeek-V4.1-Flash
Original
82.86
87.71
63.00
71.19
91.01
+RES
5.71
14.20
46.00
60.38
43.82
GPT-5.6 Sol
Original
85.71
91.24
61.00
70.60
88.76
+RES
26.67
35.44
49.00
65.08
44.94
Claude Opus 5
Original
86.67
92.13
70.00
74.98
89.89
Table 1: Environment scaling lowers task completion across six models and three benchmarks. Original denotes the released benchmark environment; +RES denotes the last accepted environment after scaling. Each model makes one 50-turn attempt per task in both conditions. All metrics use a 0–100 scale. Claw Pass@1 requires native score ≥0.80 , and Score is the mean native score. Workspace TCR@80 requires at least 80% of a task’s rubric items to pass; RPR is the percentage of passed rubric items pooled across tasks. Terminal Pass@1 uses binary verifier outcomes.
Qwen3.5-4B
Qwen3.5-9B
Qwen3.5-35B-A3B
Metric
Base
+SFT w/o RES
+SFT w/ RES
Base
+SFT w/o RES
+SFT w/ RES
Base
+SFT w/o RES
+SFT w/ RES
Claw-Eval-Live ( N=105 )
Pass@1
49.52
52.38
61.90
60.00
62.86
71.43
61.90
64.76
77.14
Score
71.51
74.12
79.15
77.75
80.32
81.04
78.96
80.36
82.83
Workspace-Bench-Lite ( N=100 )
TCR@80
1.00
4.00
11.00
9.00
12.00
21.00
11.00
14.00
28.00
Table 2: SFT with RES trajectories improves performance on the original benchmarks. For each Qwen3.5 Base model, +SFT w/o RES uses original-environment trajectories; +SFT w/ RES uses only trajectories collected in RES-scaled environments. Both use the same initialization, sample count, and optimization. Each checkpoint makes one 50-turn attempt per task. Metrics use a 0–100 scale: Claw Pass@1 thresholds native scores at 0.80, Score averages native scores, Workspace TCR@80 thresholds task rubric completion at 80%, RPR is the pooled percentage of passed rubric items, and Terminal Pass@1 uses its binary verifier. Bold marks the best result per model.
Variant
Dependency composition
Inheritance
Library
Pass@1 (%) ↓
Score ↓
A. Remove dependencies from saved environments (15 tasks)
Saved RES
✓
✓
Full
0.00
7.73
w/o dependencies
×
✓
Full
100.00
92.93
B. Rebuild environments under matched budgets (15 tasks)
Rebuilt RES
✓
✓
Full
6.67
14.40
w/o inheritance
✓
×
Full
93.33
88.73
Table 3: Effects of RES components and recursion depth on Claw-Eval-Live. DeepSeek-V4.1-Flash makes one attempt per environment with a 50-turn limit. Pass@1 uses native score ≥0.80 ; Score is mean native score on a 0–100 scale. Lower values indicate harder environments. A removes composed dependencies and their prerequisite work from saved environments. B uses the same 15 tasks with matched planned depths and proposer budgets: w/o inheritance evaluates the last independent enhancement from the original environment, and Basic retains six operators. Checkmarks mark enabled components; blue shading identifies full RES controls and the default depth cap of ten. C reevaluates 105 separate lineages, retaining the last accepted environment after early stopping.
Figure 4: Historical performance and interaction demand on Claw-Eval-Live. Lines and bars show 105-seed historical means (Equation 5 ); shading shows task interquartile ranges. The dashed line marks the 50-turn limit. Tokens include target input and output. Termination rates use logged outcomes; actual depth counts appear in Appendix C.2 .
Figure 5: Proposer cost across recursive stages. Bars and lines show the same per-child means on Claw-Eval-Live. Depths 1–4 contain 105, 94, 57, and 18 children; dark bars pool 15 children from five seeds at depths 5–10. Time sums proposer execution durations.
Figure 6: Recursive scaling of a financial reconciliation task. Access and recovery dependencies accumulate while the reconciliation objective is preserved. The diagram shows E0 – E2 ; endpoint evaluation uses final E3 .
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Environment scaling
Depth reevaluation
Model comparison
Attempts per environment
3 at evaluated nodes
1
1
Model-turn limit
60
50
50
Success criterion
Complete task fulfilment
Native score ≥0.80
Native score ≥0.80
Target
DeepSeek-V4.1-Flash
DeepSeek-V4.1-Flash
Six evaluation models
Task cohort
105 seeds
105 seeds / 394 environments
105 matched seeds per model
Primary use
Stopping decisions
Recursive trends
Original vs. RES
Appendix
Table A1: Protocols for environment scaling and independent evaluation on Claw-Eval-Live. Scaling-stage attempts determine when to stop recursion; depth reevaluation measures changes over saved lineages; the paired model comparison evaluates original and RES-scaled environments. A model turn is one target-model response, potentially containing multiple tool calls. Stopping uses complete fulfilment; independent success uses a native-score threshold of 0.80.
Quantity
Mean
Median
P90
Outer proposer executions
1.45
1.00
2.00
Input + output tokens (k)
89.19
65.27
179.82
Summed execution time (s)
112.61
93.56
184.00
Appendix
Table A2: Proposer construction cost per accepted Claw-Eval-Live child. Mean, median, and 90th-percentile (P90) costs cover all 289 children constructed by GPT-5.6 Sol with high reasoning effort. Each child’s costs aggregate its outer executions and retries. Tokens include input and output; time sums execution durations. Target solving and grading are excluded.
Setting
4B-Base
9B-Base
35B-A3B-Base
Peak learning rate
2×10−5
10−5
5×10−6
Epoch / updates
1 / 527
Maximum sequence length
49,152 tokens
Effective batch size
16 (12 in the last update)
Learning-rate schedule
Cosine, 3.00% warmup (16 steps)
Optimizer
DeepSpeedCPUAdam (AdamW)
Appendix
Table A3: Optimization settings for the Qwen3.5 Base checkpoints in the training comparison. +SFT w/o RES and +SFT w/ RES start independently from the same checkpoint and use these identical settings at each model size. Each condition trains on 8,428 trajectories for one epoch; sequence lengths and effective batch size are measured in tokens and examples, respectively. The checkpoint after the single training epoch is used for evaluation.
Figure A1 : Current retained-state performance and interaction demand. Means and interquartile ranges cover all 105 seeds, retaining early-stopped finals without averaging earlier budgets.
Figure A2 : Current versus historical benchmark means. Historical means average the retained measurements from budgets zero through the current budget.
Figure A3 : Historical interaction demand at selected budgets. Per-task means over retained states 0,…,d are averaged equally across 105 seeds, including failed and turn-limited trajectories. Tokens include repeated input context.
d
Reach d
Continue
Pass@1
Score
Turns
Tools
Tokens (M)
0
105
105
87/105
87.84
3.55
8.05
0.01
1
105
94
63/105
66.69
33.71
151.79
1.66
2
94
57
18/105
25.17
46.91
217.47
2.85
3
57
18
8/105
15.30
48.81
236.37
3.00
4
18
5
6/105
13.65
49.00
235.34
3.02
5
5
4
4/105
11.84
49.10
236.54
3.04
Appendix
Table A4: Performance and interaction cost across recursion-depth budgets on Claw-Eval-Live. DeepSeek-V4.1-Flash receives one 50-turn attempt per environment. Each budget d covers 105 tasks, retaining stopped environments. Reach d counts chains attaining depth d ; Continue counts chains exceeding it. Pass@1 counts scores ≥0.80 ; Score is the native mean on a 0–100 scale. Turns, Tools, and Tokens are task means; tokens sum target input and output in millions.
Figure A4 : Performance across recursion-depth caps. Pass@1 (orange) and mean native score (blue) use a 0–100 scale over 105 seeds, retaining stopped environments. Table 3 selects caps 0, 1, 2, 3, 5, and 10.
Task
di
Original
Saved
w/o deps.
Rebuilt
w/o rec.
Basic
A04
10
93.00
0.00
89.00
0.00
97.00
95.00
D02
4
96.00
0.00
96.00
98.00
99.00
98.00
COMM_13
5
100.00
4.00
82.00
3.00
100.00
82.00
DATA_09
8
86.00
7.00
93.00
3.00
88.00
88.00
PRODAPP_01
3
92.00
14.00
99.00
15.00
99.00
97.00
DATA_07
3
96.00
0.00
96.00
3.00
100.00
98.00
Appendix
Table A5: Per-task native scores for the 15-task component ablations. DeepSeek-V4.1-Flash receives one 50-turn attempt per environment; scores use a 0–100 scale, with success at 80 or above. Original is the unmodified task. Saved and Rebuilt denote full RES controls; w/o deps. removes dependencies from Saved, w/o rec. independently enhances Original and evaluates the predeclared last construction, and Basic rebuilds with six operators. The rebuilt variants share planned depth di .
Figure A5 : Per-task effects of the component interventions. Filled blue markers show the corresponding full RES control; hollow orange markers show the ablation. Segments connect the same task, and dashed lines mark the 80-point success threshold. All panels use the task order in Table A5 .
Condition
Turns
Tool calls
Tokens (k)
Relative tokens ( × )
Original
2.87
6.67
8.43
1.00
Saved RES
50.00
226.93
3115.42
369.40
w/o dependencies
19.07
85.60
723.76
85.82
Rebuilt RES
49.73
163.93
1644.20
194.96
w/o inheritance
21.20
90.40
651.09
77.20
Basic library
18.53
39.60
227.01
26.92
Appendix
Table A6: Target execution cost for the 15-task component comparisons. Values are task means for DeepSeek-V4.1-Flash with one 50-turn attempt per environment, including failed and turn-limited attempts. Saved RES is the dependency-removal control; Rebuilt RES is shared by the independent-construction and basic-library comparisons. Tokens sum target input and output; k denotes thousands. Relative tokens divides the condition’s mean by the Original mean before rounding. Proposer construction and grading costs are excluded.
Condition
τ=0.70
τ=0.80
τ=0.90
Original
15/15
15/15
13/15
Saved RES
0/15
0/15
0/15
w/o dependencies
15/15
15/15
10/15
Rebuilt RES
1/15
1/15
1/15
w/o inheritance
14/15
14/15
12/15
Basic library
14/15
12/15
10/15
Appendix
Table A7 : Success counts under alternative native-score thresholds on Claw-Eval-Live. Each entry counts tasks with score ≥τ over the cohort. Component rows use 15 paired tasks; depth rows use 105 reevaluated seeds with stopped environments retained. Counts use saved scores.
Generalizable agents should adapt to diverse tasks and unseen environments beyond their training distribution. This position paper argues that such generalization requires environment scaling: expanding the distribution of executable rule-sets that agents interact with, rather than only increasing trajectories or tasks within fixed benchmarks. Current scaling practices largely focus on collecting more experience or broader task sets under fixed interaction rules, leaving agents brittle when underlying interfaces, dynamics, observations, or feedback signals change. The core challenge is therefore a world-level distribution shift: agents need systematic exposure to environments with meaningfully different executable rule-sets. To clarify this challenge, we propose a unified taxonomy that separates trajectory scaling, task scaling, and environment scaling by their primary deliverables and by what changes in the executable rule-set. Building on this taxonomy, we synthesize construction paradigms for scalable environments, contrasting programmatic generators that prioritize controllability and verifiability with generative world models that offer broader coverage and open-endedness. We further outline how environment scaling can be coupled with stateful learning mechanisms, emphasizing learned update rules for cross-environment adaptation. We conclude by discussing alternative perspectives and argue that scalable environments provide the essential substrate for measurable and controllable progress toward robust general agents.
Existing Agent benchmarks suffer from two critical limitations: high environment interaction overhead (up to 41% of total evaluation time) and imbalanced task horizon and difficulty distributions that make aggregate scores unreliable. To address these issues, we propose AgentCE-Bench built around a unified grid-based planning task, where agents must fill hidden slots in a partially completed schedule subject to both local slot constraints and global constraints. Our benchmark offers fine-grained control through two orthogonal axes: Scalable Horizons, controlled by the number of hidden slots H, and Controllable Difficulty, governed by a decoy budget B that determines the number of globally misleading decoy candidates. Crucially, all tool calls are resolved via static JSON files under a Lightweight Environment design, eliminating setup overhead and enabling fast, reproducible evaluation suitable for training-time validation. We first validate that H and B provide reliable control over task horizon and difficulty, and that AgentCE-Bench exhibits strong domain consistency and model discriminability. We then conduct comprehensive experiments across 13 models of diverse sizes and families over 6 domains, revealing significant cross-model performance variation and confirming that AgentCE-Bench provides interpretable and controllable evaluation of agent reasoning.
Wang Yang, Chaoda Song, Xinpeng Li +7
Case Western Reserve University · NII LLMC (Japan) · University of Zurich
Terminal agents solve tasks in software engineering and scientific research by executing commands in a terminal. Training them via reinforcement learning (RL) requires diverse and verifiable environments, but these environments are still mostly designed by human researchers, so their supply grows only as fast as researchers can work. We introduce AutoEnvScaling, which treats environment design as a terminal task and uses agents to automate the data flywheel. A proposer agent builds new RL environments from the rollouts of the solver being trained, and the solver's new rollouts guide the next round. By turning environment generation into a terminal task, AutoEnvScaling increases the valid-environment rate by up to 25 times compared to prompt-only generation, at up to 75% lower cost per valid environment. We show that training Qwen3.5-35B-A3B on AutoEnvScaling environments from GPT-5.6-sol improves its Terminal-Bench 2.1 score by 4.3 points over the matched Tmax baseline. The broader goal of AutoEnvScaling is to enable recursive self-improvement (RSI): by making both environment generation and solving terminal tasks, it lets one model co-evolve in both roles. We show that Qwen3.6-35B-A3B, acting in both roles, gains 12.2 points on Terminal-Bench 2.1 and raises its Terminal-Bench 4.0 score from 1.6 to 3.2. The pipeline also generalizes across agentic domains, generating environments for software engineering, computer use, professional work, and GPU kernel programming. Overall, these results show a path toward agents building their own training environments, reducing human reliance and taking a core step toward RSI.
Simon Yu, Baolin Peng, Bo Liu +8
Northeastern University · Microsoft · Stanford University +2