acceptodds
Under review as a conference paper at ICLR 2027

AgentMercury: Your Agent Can Synthesize Verifiable Environments for Business Scenarios at scale

Abstract

Agents learn to act through interaction with environments, yet the environments used for training are often manually constructed or synthesized around predefined tasks and benchmarks. This task-centric paradigm makes it difficult to scale environments that reflect realistic and evolving workflows where diverse tasks can naturally emerge from the underlying world. We introduce AgentMercury, a scalable framework for synthesizing executable environments from high-level business scenarios. Rather than constructing an environment for a specific task, AgentMercury first instantiates a persistent world with entities, services, tools, state, and executable cross-service invariants, from which diverse tasks and interaction trajectories can subsequently emerge. We construct 4,783 executable environments spanning 14 industries and 50 countries, and use them as training substrates for reinforcement learning. Despite being generated without targeting the evaluation benchmarks, policies trained on these business-oriented environments improve substantially on both enterprise workflows and out-of-domain benchmarks spanning reasoning, coding, scientific computing, and tool use. In our experiments, Qwen3.5-4B improves from 12.3 to 15.7 on EnterpriseOps-GYM and from 45.9 to 56.0 on AIME26 after training on AgentMercury environments. We further show that the construction process itself can be learned: fine-tuning Qwen3.5-35B-A3B on construction traces increases executable-world authoring success from 3.3% to 83.3% on held-out business scenarios. These results show that scenario-grounded environments can provide useful and generalizable learning signals beyond benchmark-specific training, while their construction can itself become a learnable capability.

Primary area:
reinforcement learning
Keywords:
Environment Synthesis, Task Synthesis, MCP

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_SZFHOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces AgentMercury, a framework that first builds executable environments from business scenarios and then generates tasks from those environments. The authors develop PLANET to construct worlds with multiple services, tools, persistent state, and cross-service constraints. They generate 4,783 environments and use the resulting tasks to train Qwen3.5 models with RL, reporting improvements on enterprise workflows and several out-of-domain benchmarks.

Critical strengths

I think the most interesting part of this paper is its world-first approach. Many existing methods build environments around specific tasks, whereas AgentMercury starts with a persistent business world where different services interact and share state, and then generates tasks from it. This makes the environment less tied to any single predefined task and allows the same world to support different kinds of interactions. I see this as a worthwhile direction, especially for agents that need to handle long workflows across multiple systems.

Critical weaknesses

My main concern is that the paper does not really establish the value of its world-first design. The experiments show that RL on the generated tasks improves model performance, but they don't tell us whether the improvement comes from the world-first approach itself. Would a task-first or another environment generation method achieve similar results under the same training budget? I think this is the most important question the paper needs to address. The framework also combines world reuse, multi-service structure, persistent state, and cross-service constraints, without ablations that separate their effects. Even with the positive results, it is hard to tell which of these design choices actually matter.

Some experimental results also raise questions. On τ³ Telecom, the 35B base model scores 49.1 ± 49.8, while the 4B base model reaches 92.5. Such large variance makes me less convinced by the authors' claims about policy stability. The 4B SAO results are also omitted due to unstable training, which I think is worth reporting.

Finally, the environment-authoring results improve from 3.3% to 83.3%, but the evaluation uses only 30 held-out scenarios. The success rate also drops to 10% with recipe conditioning. The authors attribute this to the model having internalized the construction process, but it could just as well reflect sensitivity to the input format. The current experiments do not distinguish between these explanations.

Questions for authors
  1. Could the authors compare AgentMercury with a task-first or other environment generation approach under the same RL training budget? I think this comparison is important for understanding whether the gains actually come from the world-first design. Ablations on world reuse, persistent state, and cross-service constraints would also help clarify which parts of the framework matter most.

  2. Could the authors explain the unusually large variance on τ³ Telecom, particularly for the 35B base model? I would also like to see the omitted 4B SAO results, even if training was unstable.

  3. What evidence supports the authors' explanation for the drop from 83.3% to 10% with recipe conditioning? Could this simply reflect sensitivity to the input format rather than interference with an internalized construction process?

6 comments
AuthorsOpenReview ↗

Why construct the world first, and what the current evidence establishes

Thank you for recognizing the world-first direction and for distinguishing the utility of the generated training data from evidence about individual design choices.

W1 and Q1. World-first design, matched-budget alternatives, and component ablations.

My main concern is that the paper does not really establish the value of its world-first design. The experiments show that RL on the generated tasks improves model performance, but they don't tell us whether the improvement comes from the world-first approach itself. Would a task-first or another environment generation method achieve similar results under the same training budget? I think this is the most important question the paper needs to address. The framework also combines world reuse, multi-service structure, persistent state, and cross-service constraints, without ablations that separate their effects. Even with the positive results, it is hard to tell which of these design choices actually matter.

We have not completed a matched-RL-budget comparison with another environment source or component-removal ablations. The reported improvements establish the training utility of the complete pipeline; they do not show that world-first construction causes those improvements or outperforms a task-first source. The Introduction and Section 4.1 now make that distinction explicit.

The reason for constructing a world first is to give task synthesis a reusable set of business entities, ownership relations, and operations. The composer selects owned records in that world and attaches situation-specific evidence, instructions, and grading criteria. Appendix A illustrates this with two released portfolio investigations that reuse the same customer record and tool interface. One attaches incident-and-coverage evidence; the other attaches quality-issue evidence. Their complete investigation structure is shared after entity/value rebinding. This is evidence of reuse across situations, not of distinct whole-task objective families or superior training efficiency.

The components serve different purposes: persistent state retains action consequences within an episode; services expose records and communication channels; cross-service constraints specify consistency requirements. These implemented roles explain the design, but do not establish each component's contribution to learning. We also distinguish declaring a constraint from enforcing it during every transition.

Appendix A provides two supporting analyses within that scope. Its construction-cost identity explains how compatible tasks can share world-building overhead; AWM and Agent-World also reuse environments, so this is not an asymptotic or measured cost advantage over them. Its scripted execution example verifies one request-grounded state condition using native tools, without establishing full-task success or an RL effect. The requested comparative learning evidence remains a substantive limitation.

W2 and Q2. Telecom variance and the omitted 4B SAO results.

Some experimental results also raise questions. On τ³ Telecom, the 35B base model scores 49.1 ± 49.8, while the 4B base model reaches 92.5. Such large variance makes me less convinced by the authors' claims about policy stability. The 4B SAO results are also omitted due to unstable training, which I think is worth reporting.

We agree that the unusually large variance on τ³ Telecom warrants further discussion. Table 2 reports 49.1 ± 49.8 for the 35B base model over three evaluation runs, compared with 92.5 ± 1.0 for the 4B base model. Such variability is not necessarily unique to our setting, as instability and sensitivity in base-model performance have also been discussed in prior studies and external reports. Nevertheless, without the original run-level outputs and serving configurations, we cannot determine the specific cause of the observed variance.

Interestingly, both RL training methods show a trend toward reduced variability relative to the 35B baseline, with the standard deviation decreasing from 49.8 to 20.5 (GRPO) and 23.8 (SAO). While these results suggest that RL training may help mitigate some of the underlying variability, the remaining variance is still substantial, and three evaluation runs are insufficient to establish a general stabilization effect. We have therefore revised Section 4.3 and the Conclusion to highlight this observed trend without making a strong causal claim about policy stability.

We also agree that the excluded 4B SAO outcome should be reported. Unfortunately, validated benchmark values are not available in the recovered materials, so we cannot provide reliable numerical results or infer them from the training curve. Appendix E.3 and the Conclusion now explicitly acknowledge the exclusion and the reported training instability. Appendix C also clarifies that the optimizer comparison uses broadly comparable infrastructure but does not isolate the optimizer's causal effect.

W3 and Q3. Authoring sample size and why the recipe hurts.

Finally, the environment-authoring results improve from 3.3% to 83.3%, but the evaluation uses only 30 held-out scenarios. The success rate also drops to 10% with recipe conditioning. The authors attribute this to the model having internalized the construction process, but it could just as well reflect sensitivity to the input format. The current experiments do not distinguish between these explanations.

Qwen3.5-35B-A3B authoringBrief onlyBrief + recipeUpto 200 instances (recipe)
Before construction-trace fine-tuning3.3%20.0%77.5%
Digest only10.0%76.7%67.0%
Exemplar only16.7%73.3%71.0%
Digest + Exemplar20.0%10.0%25.5%
After fine-tuning83.3%10.0%65.5%

The recipe changes success in opposite directions before and after fine-tuning. Each condition evaluates 30 held-out briefs, with success requiring all 12 structural validators (Table 3; Appendix F). The paper reports that 27 fine-tuned recipe-conditioned outputs fail the cross-service check. This identifies an unmet structural requirement in the outputs, without identifying how it arose. The recipe jointly adds an invariant digest and an exemplar, changing both content and input format. Format sensitivity, exemplar content, effective generation headroom, and interference with a learned procedure are not separated by this comparison.

We have therefore withdrawn interference with an internalized construction process as an established explanation and retained it only as a hypothesis. The revised authoring claim is limited to the tested brief-only interface. The recipe-conditioned failure is established; its cause and generality remain unresolved. Distinguishing these explanations would require paired tests of the original checkpoint with digest-only, exemplar-only, and combined inputs under equal effective output budgets.

Dear Reviewer,

We sincerely appreciate your time and thoughtful feedback on our work.

We understand that this is a particularly busy period and that reviewing additional materials can be demanding. Nevertheless, we would be very grateful if you could take another look at our responses and revised manuscript, where we have made substantial efforts to address your concerns.

We hope these revisions help clarify our contributions and resolve the issues you raised. If you find that our responses and revisions adequately address your concerns, we would sincerely appreciate your reconsideration of the overall assessment and score.

Thank you again for your valuable time and consideration.

Reviewer_DHcxOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

AgentMercury generates executable business environments from high-level scenarios, then instantiates tasks within them. The framework constructs persistent state, tools, and cross-service invariants. The authors generate 4,783 environments and train Qwen models on the trainable subset, obtaining gains on enterprise workflows and external reasoning and tool-use benchmarks.

Critical strengths
  1. Separating world construction from task generation supports reusable environments and cross-service workflows, which enhances the quality of the synthesized environments.
  2. Training gains extend beyond the generated tasks to external benchmarks and two model scales, which fully justify the motivation of this work.
Critical weaknesses
  1. Authors should further justify the sourcing seeds behind this large-scale synthesis and how they support the synthesized task scenarios, tool environments, and verifiable rubrics.
  2. It seems interesting that the enterprise distribution dataset generalizes very well to OOD tasks like math and code. The transferability mechanism behind it should be further explained and justified with the corresponding ablations.
  3. Adding the construction recipe reduces success from 83.3% to 10.0%. This large drop raises concerns about robustness beyond the pipeline.
Questions for authors

See weakness.

3 comments
AuthorsOpenReview ↗

Business-grounded seeds, broader evaluation, and the scope of recipe robustness

Thank you for recognizing the value of separating world construction from task generation and for asking how the seeds, external gains, and authoring robustness are supported.

W1. Seed sources and their connection to scenarios, tools, and rubrics.

Authors should further justify the sourcing seeds behind this large-scale synthesis and how they support the synthesized task scenarios, tool environments, and verifiable rubrics.

The seeds were profiles of leading real companies, collected by country from public Wikipedia snapshots in June and July 2026. They supplied company-level business context, including industry, scale, and basic attributes. The operational records, employees, and communications in the worlds were generated, and company-identifying information was anonymized. The collection window describes when the sources were consulted, not a common fiscal year for every attribute.

The pipeline connects that context to verifiable task conditions in stages. PLANET uses the profile to construct the fictional business's services, schema, tools, seeded state, and cross-service constraints (Section 3.1). The task composer then selects a persona and owned records, assigns situations, and creates the associated messages, intended actions, and rubric requirements (Section 3.2). Grading evaluates the resulting interaction against the task-specific specification. Thus, task requirements refer to generated records and evidence in an executable world, rather than factual answers about the real company.

The profiles provide contextual grounding for synthesis; they do not validate the fictional workflows as faithful reconstructions of a real company. Section 3.1, the Figure 1 caption, and the Ethics Statement now distinguish collected business context from generated operational content. Appendix A separately illustrates how a generated task refers to executable state: an original request and rubric clause require resolving an account issue, and scripted tool calls produce that state change. This example connects a task requirement to a state condition; it does not establish a historical trace from a collected company attribute to that task.

W2. Why mathematics and coding improve, and the requested mechanism ablation.

It seems interesting that the enterprise distribution dataset generalizes very well to OOD tasks like math and code. The transferability mechanism behind it should be further explained and justified with the corresponding ablations.

We observe the gains, but have not run an ablation that identifies their cause. The external evaluations were initially intended to check whether business-workflow training damages broader capabilities. They evaluate the complete pipeline, including environment construction, task synthesis, grading, and policy learning. Table 2 reports:

ModelAIME26HMMTLiveCodeBenchSciCode
Qwen3.5-4B45.9 ± 1.728.5 ± 1.236.6 ± 2.822.6 ± 0.1
+ GRPO + AgentMercury56.0 ± 1.835.4 ± 1.644.0 ± 1.225.7 ± 0.7
Qwen3.5-35B-A3B91.0 ± 0.377.0 ± 0.974.3 ± 2.229.7 ± 1.5
+ GRPO + AgentMercury91.9 ± 2.180.0 ± 0.779.0 ± 0.529.9 ± 1.7
+ SAO + AgentMercury92.2 ± 1.983.3 ± 3.178.6 ± 1.028.3 ± 1.9

Values are means ± standard deviations over three evaluation runs. SciCode reports subproblem accuracy. Across all nine benchmarks and three base-to-trained conditions in Table 2, 25 of 27 mean-score comparisons increase. This is a descriptive count, not independent replications or a significance test. The exceptions are 4B Telecom (92.5 to 91.9) and 35B SAO SciCode (29.7 to 28.3); several other changes are small relative to run-to-run variation. We therefore do not claim universal capability preservation or formal noninferiority.

A candidate explanation is transfer of procedures for selecting evidence and controlling state changes. The released rubrics require following dated evidence, integrating reports across channels, and withholding a record change while reporting a conflicting claim. These requirements motivate the hypothesis, but do not explain the mathematics and coding gains: task content alone shows neither which procedures the policy learned nor whether those procedures caused transfer. Section 4.3 and the Conclusion now separate the empirical utility of the integrated training pipeline from this untested explanation.

W3. The authoring drop with recipe conditioning.

Adding the construction recipe reduces success from 83.3% to 10.0%. This large drop raises concerns about robustness beyond the pipeline. The decline exposes a substantial robustness limitation under the tested recipe; its cause remains unidentified:

Qwen3.5-35B-A3B authoringBrief onlyBrief + recipeUpto 200 instances (recipe)
Before construction-trace fine-tuning3.3%20.0%77.5%
Digest only10.0%76.7%67.0%
Exemplar only16.7%73.3%71.0%
Digest + Exemplar20.0%10.0%25.5%
After fine-tuning83.3%10.0%65.5%

The recipe changes success in opposite directions before and after fine-tuning. Each condition evaluates 30 held-out briefs, with success requiring all 12 structural validators (Table 3; Appendix F). The paper reports that 27 fine-tuned recipe-conditioned outputs fail the cross-service check. This identifies an unmet structural requirement in the outputs, without identifying how it arose. The recipe jointly adds an invariant digest and an exemplar, changing both content and input format. Format sensitivity, exemplar content, effective generation headroom, and interference with a learned procedure are not separated by this comparison.

We have therefore withdrawn interference with an internalized construction process as an established explanation and retained it only as a hypothesis. The revised authoring claim is limited to the tested brief-only interface. The recipe-conditioned failure is established; its cause and generality remain unresolved. Distinguishing these explanations would require paired tests of the original checkpoint with digest-only, exemplar-only, and combined inputs under equal effective output budgets.

Dear Reviewer,

We sincerely appreciate your time and thoughtful feedback on our work.

We understand that this is a particularly busy period and that reviewing additional materials can be demanding. Nevertheless, we would be very grateful if you could take another look at our responses and revised manuscript, where we have made substantial efforts to address your concerns.

We hope these revisions help clarify our contributions and resolve the issues you raised. If you find that our responses and revisions adequately address your concerns, we would sincerely appreciate your reconsideration of the overall assessment and score.

Thank you again for your valuable time and consideration.

Reviewer_KDZgOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AgentMercury, a framework that constructs executable business worlds from high-level scenarios, then generates specific tasks. AgentMercury synthesized 4000+ environments across 14 industries and 50 countries. Training on those synthesized environments improves performance on enterprise benchmarks (val set of AgentMercury) and several agent benchmarks. This paper also studies the ability of LLMs to create new environments by fine-tuning on construction traces, which improves Qwen3.5-35B-A3B's performance. I find that many appendix content is mentioned in the main text, but actually missing in the Appendix (only 14 pages of the uploaded PDF).

Critical strengths
  1. The synthesized resource covers 4000+ environments across 14 industries and 50 countries, which may be useful for future research, plus the synthesis used open-source model DeepSeek-v4-flash.
  2. After fine-tuning on these environments, the performance of Qwen models on OOD benchmarks improves, indicating the utility of the synthesized environments for training agents.
  3. This paper also studies the ability of LLMs to create new environments by fine-tuning on construction traces, which improves Qwen3.5-35B-A3B's performance. This offers a promise for future self-improving loop, where LLMs can generate new environments and improve themselves through fine-tuning.
Critical weaknesses
  1. The presentation needs to be improved. For example, the main figure (Figure 2) is not clear (I do not against AI generated figures, but the figure should be clear and readable). The font in Figure 3 is too small. Many mentioned Appendix content is missing in the uploaded PDF (only 14 pages), please carefully check the paper before submission.
  2. The main experiments only report the performance of models before and after AgentMercury training, without a matched-budget alternative environment source or ablations of synthesis stages.
  3. There is a lack in analysis of synthesized environment & task quality, for example, the diversity, the functionality, and the task difficulty. It would be better to provide more analysis on the data itself.
  4. The synthesis pipeline seems to be not new, can the authors provide more discussion on the novelty of the synthesis pipeline compared to existing works, which can be enhanced in Related Work section.
Questions for authors

Please see the weakness.

5 comments
AuthorsOpenReview ↗

Reusable business worlds, task-level evidence, and a complete revised manuscript

Thank you for the reviewer's constructive feedback and for identifying the need to connect the environment resource to the tasks it supports.

W1. Figure readability and the missing appendix.

The presentation needs to be improved. For example, the main figure (Figure 2) is not clear (I do not against AI generated figures, but the figure should be clear and readable). The font in Figure 3 is too small. Many mentioned Appendix content is missing in the uploaded PDF (only 14 pages), please carefully check the paper before submission.

At first, we included our appendix in the supplementary details and tried to update the manuscript regarding your concern. We have redrawn Figure 2 around four sequential phases: world construction, task instantiation, policy training, and evaluation. Each phase shows its inputs and outputs, including the separation between the instruction seen by the policy and the hidden rubric used by the grader. Figure 3 retains the horizontal layout and original curves, with larger vector labels and all checkpoint and score ticks visible. The revised manuscript contains Appendices A–F, covering world/task statistics, the interaction model, training settings, benchmark protocols, policy-learning results, and authoring details.

W2. Matched-budget alternatives and synthesis-stage ablations.

The main experiments only report the performance of models before and after AgentMercury training, without a matched-budget alternative environment source or ablations of synthesis stages.

We have not completed the matched-budget source comparison or synthesis-stage ablations you request. The before/after results evaluate the complete training pipeline, so they cannot attribute the gains specifically to world-first construction or to an individual stage. Section 4.1 now states this limit. EnterpriseOps-Gym is an external benchmark rather than an AgentMercury validation split; this clarifies the evaluation's scope but does not substitute for the requested comparison.

World-first construction addresses a design requirement: establish records, ownership, and operations that task synthesis can reuse. The added Appendix A example follows the same customer record and tool interface across two portfolio investigations. One binds incident-and-coverage evidence to that record; the other binds quality-issue evidence. Their full investigation structure is shared. This demonstrates reuse across situations, while leaving comparative training effectiveness open.

W3. Environment coverage, task content, functionality, and difficulty.

There is a lack in analysis of synthesized environment & task quality, for example, the diversity, the functionality, and the task difficulty. It would be better to provide more analysis on the data itself.

We added coverage and task-content statistics and a bounded execution example. We have not measured task difficulty for a specified policy. The revision connects the reported library of 4,783 environments, 14 industries, and 50 countries to explicit task statistics. Appendix A analyzes the released 43,300-task slice, including its TRAIN and TEST splits, separately from the full construction library:

Rubric requirementTasksShare of the released slice
Post to a communication channel40,54393.6%
Send an email27,44463.4%
Preserve a record despite a conflicting claim25,80759.6%

These varying requirements sit within a common investigation structure: every released rubric requires retrieval, updates, a summary, and preservation of specified records. The categories overlap and describe task specifications, not observed behavior or independent whole-task objective families. The new action–target figure summarizes 799,513 classified action-bearing rubric lines, with separate denominators for lines and tasks. A companion figure shows industry coverage, exposed services/tools, rubric length, and within-world situation combinations. Across 4,307 unambiguously identified source-world groups, the ten-seed slice has a median of eight situation-label sets per world. These are generator configurations, not independently verified semantic goals.

We also added a bounded execution example. In an unchanged portfolio task, an older account message explicitly marks itself superseded; a later message confirms resolution. Scripted calls to the released runtime read the evidence and update the account. We derived a completion predicate from the original request and its rubric clause and applied it to the native final state:

Scripted actionFinal target statusDerived predicate
Update the identified accountresolvedPass
Make no changeopenFail
Update a different accountopenFail
Write an incorrect statuspendingFail

An independent replay reproduced these outcomes. The predicate is separate from the task's released state assertions. This validates one request-grounded state condition, not whole-task completion or model success. Rubric length, situation counts, and this scripted replay do not measure policy-conditional difficulty. Appendix A makes these units and boundaries explicit.

W4. Novelty relative to existing synthesis.

The synthesis pipeline seems to be not new, can the authors provide more discussion on the novelty of the synthesis pipeline compared to existing works, which can be enhanced in Related Work section.

We have revised the Introduction and related works to explain the implemented business-profile-to-training path: business context determines services, state, and consistency requirements; task instantiation connects selected records to evidence and grading rubrics; the interaction interface supplies trajectories for policy learning. Construction traces additionally support learning to author worlds, evaluated separately from learning to act in them.

We do not claim that reuse or environment-first ordering alone is unique. AWM conditions a shared environment on an initial task set, while Agent-World constructs environments before synthesizing tasks from their tool structure; both reuse environments. Appendix A accordingly presents shared construction cost as a conditional amortization rationale, not a measured cost or asymptotic advantage. The contribution we defend is the integrated business-grounded pipeline, the resource it produces, and the reported training utility of that complete system.

Dear Reviewer,

We sincerely appreciate your time and thoughtful feedback on our work.

We understand that this is a particularly busy period and that reviewing additional materials can be demanding. Nevertheless, we would be very grateful if you could take another look at our responses and revised manuscript, where we have made substantial efforts to address your concerns.

We hope these revisions help clarify our contributions and resolve the issues you raised. If you find that our responses and revisions adequately address your concerns, we would sincerely appreciate your reconsideration of the overall assessment and score.

Thank you again for your valuable time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Overall Response to Reviewers

We would like to sincerely thank the Program Chairs, Senior Area Chairs, Area Chairs, and Reviewers for their time and efforts in organizing and conducting the review process, and for giving us the opportunity to present and improve our work. We are particularly grateful to the reviewers for their thoughtful comments, constructive suggestions, and careful evaluation of our manuscript. Their feedback has been invaluable in helping us identify areas for improvement and strengthen the overall quality of our work.

In response to the reviewers' comments, we have carefully revised the manuscript to provide additional empirical evidence, clarify our main contributions, and better acknowledge the limitations of our current study.

The key revisions are summarized below:

1. Environment and Task Quality: We have expanded our analysis of the 4,783 constructed environments and 43,300 released tasks, providing additional details on industry coverage, task requirements, and cross-service interactions. We have also included an executable state-validation example to better illustrate the functionality of the generated environments.

2. World-First Design and Novelty: We have further clarified the motivation and design choices behind our world-first approach, including concrete examples demonstrating how business environments, records, and tools can be reused across different task situations. We have also expanded the discussion of related work to better position our contributions relative to existing environment-generation approaches.

3. Training Effectiveness and Generalization: We have provided further clarification of the evaluation results on external benchmarks, where improvements were observed in 25 out of 27 mean-score comparisons. We have also carefully distinguished these empirical observations from potential explanations that would require additional experiments to validate.

4. Authoring Robustness and Experimental Reliability: We have expanded our discussion of recipe-conditioning experiments, the observed variance in Telecom evaluations, and the excluded 4B SAO results. Based on the reviewers' suggestions, we have also refined our interpretations of policy stability and construction-process internalization to avoid claims that are not sufficiently supported by the available evidence.

5. Presentation and Reproducibility: We have redesigned the main pipeline figure, improved the readability of other figures, and expanded Appendices A–F with additional implementation details, experimental settings, and evaluation protocols. We hope these revisions make the manuscript clearer, more accessible, and easier to follow and reproduce.

Our central contribution is an integrated framework for constructing reusable, executable business environments and generating verifiable tasks for agent training. Through these revisions, we have aimed to provide a clearer understanding of the framework's design, the characteristics of the generated resources, and their demonstrated utility for agent training. At the same time, we acknowledge that further matched-budget comparisons and component-level ablations would be valuable for better understanding the contributions of individual design choices.

Once again, we sincerely appreciate the time, effort, and thoughtful feedback provided by the reviewers and the organizing committee. We hope that our revisions and detailed responses adequately address the major concerns and help clarify the contributions and significance of AgentMercury.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.