acceptodds
Under review as a conference paper at ICLR 2027

Evolving Multi-Agent Systems through Sense of Agency

Abstract

Multi-agent systems (MAS) tackle complex tasks by coordinating agents with complementary roles and capabilities. Evolving their individual behaviors and interaction structures during inference enables adaptation to task-specific demands, but requires informative feedback. Existing approaches often rely on external feedback from task-specific verification environments or separate LLM evaluators, limiting evolution when such feedback is unavailable. We propose SOAP, the Sense of Agency Paradigm, which draws inspiration from people's Sense of Agency (SoA)—the feeling that their own actions shape what happens. SOAP turns this intuition into internal feedback: agents anticipate what their actions will achieve and assess whether execution meets those expectations. We formulate computational SoA to capture both an agent's expectation of performing well and how closely its post-execution self-assessment matches that expectation. At individual (i.e., intra-agent) and interaction (i.e., inter-agent) levels, this feedback guides iterative prompt and topology updates toward higher anticipated achievement and smaller prediction–outcome gaps, without external feedback. Experiments on BrowseComp-Plus, Plancraft, WorkBench, and Cloudcast demonstrate superior performance against evaluated baselines, with gains across backbone models. Further analysis highlights the complementary benefits of individual- and interaction-level feedback, supporting SoA as an effective internal signal for improving task completion during MAS evolution. Supplementary materials are available [here](https://anonymous.4open.science/r/SOAP-ICLR2027-supplementary-submission), and a demo video can be downloaded [here](https://anonymous.4open.science/api/repo/SOAP-ICLR2027-supplementary-submission/file/SOAP_demo_video.mp4?v=f637803c&download=true).

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Sense of Agency, Multi-Agent System, MAS Evolution

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

This paper studies how to evolve multi-agent systems (MAS) at inference time without relying on external rewards or LLM evaluators. The authors propose SOAP, a framework inspired by the human sense of agency (SoA), which uses agents' predictions of their performance and their post-execution self-assessments to construct internal feedback signals. SOAP leverages two types of signals: individual- and interaction-level SoA to iteratively optimize agent prompts and communication topology. Experiments across four benchmarks demonstrate improvements over the evaluated baselines. Compared with prior MAS evolution methods that often rely on external feedback, the main contribution is the introduction of SoA-based internal feedback to guide test-time adaptation.

Critical strengths
  1. The idea of drawing inspiration from the human sense of agency to design internal feedback signals for MAS evolution is interesting. Exploring how agents can assess and improve their own behavior without external evaluators is a meaningful research direction.

  2. Extensive empirical evaluation. The paper evaluates the proposed approach across four benchmarks covering different tasks and several LLM backbones. The authors also provide ablation studies to investigate the contributions of individual- and interaction-level SoA.

Critical weaknesses
  1. While the human-inspired motivation is interesting, it remains unclear why the specific computational formulation of SoA is more informative than conventional self-assessment signals. In particular, the proposed score (A=p(1-|p-o|)) combines anticipated performance with prediction–outcome agreement, but the justification for this specific combination is limited. For a fixed outcome (o), this score is maximized at (p=(1+o)/2), rather than (p=o), which could favor optimistic predictions over calibrated ones. A clearer discussion of this formulation and its advantages would strengthen the methodological contribution.

  2. Although the paper includes SelfOrg (Tastan et al., ICLR 2026) and MetaGen (Wang et al., 2026) as internal-feedback baselines, these approaches involve different MAS evolution mechanisms and do not directly isolate the advantages of the proposed SoA signal over other forms of internal feedback. Existing approaches such as Reflexion and Self-Refine also demonstrate the effectiveness of self-assessment and self-reflection for iterative improvement. I would encourage the authors to consider additional controlled ablations within the same MAS evolution framework, replacing SoA with simpler internal feedback signals, such as (i) post-execution self-assessment alone, (ii) prediction–outcome discrepancy alone, or (iii) trajectory-based self-reflection without SoA. Such experiments would help establish whether the proposed SoA formulation provides additional benefits beyond standard internal feedback mechanisms.

Questions for authors
  1. Contribution of SoA versus trajectory-based self-reflection. In Section 3.4.1 and Appendix A.2 (Template 4), the evolution signal appears to be generated by reflecting on execution records, tool observations, predictions, assessed outcomes, and SoA scores. Therefore, the proposed framework seems to leverage not only SoA but also substantial trajectory-based self-reflection information. Could the authors clarify the respective contributions of these two mechanisms? In particular, would the method still achieve similar improvements if the agents performed self-reflection using the same execution trajectories and outcome assessments, but without the proposed SoA scores?

  2. Interpretation of interaction-level SoA. The interaction-level SoA compares a sender's prediction of a downstream agent's performance with the receiver's self-assessed outcome. However, this does not necessarily measure the sender's actual contribution to the receiver's performance. Could the authors clarify why this metric is an appropriate measure of interaction-level agency?

9 comments
AuthorsOpenReview ↗

Response to Weakness 1

Thank you for your review. We provide further clarification and additional results below to address your concerns. Please let us know if we have misunderstood any comments, and we will respond promptly.


Weakness 1: While the human-inspired motivation is interesting, it remains unclear why the specific computational formulation of SoA is more informative than conventional self-assessment signals. In particular, the proposed score (A=p(1−∣p−o∣)A = p(1 - \lvert p - o\rvert)) combines anticipated performance with prediction–outcome agreement, but the justification for this specific combination is limited. For a fixed outcome (oo), this score is maximized at (p=(1+o)/2p = (1 + o)/2), rather than (p=op = o), which could favor optimistic predictions over calibrated ones. A clearer discussion of this formulation and its advantages would strengthen the methodological contribution.

Response: Thank you for your valuable comments. Your observation is correct: for a fixed oo, the SoA score is maximized at p∗=(1+o)/2p^* = (1 + o)/2. We deliberately design the score to prefer predictions somewhat higher than the assessed outcome, and we explain the formulation below.

(1) Why the two terms are combined. The agreement term 1−∣p−o∣1 - \lvert p - o\rvert corresponds to the comparator cue in agency research, i.e., the match between predicted and obtained outcomes. Used alone, it is satisfied by an agent that expects little and achieves little, so it places no pressure on performance. The expectation term pp supplies that pressure, but alone it would reward unconstrained optimism. Their product rewards high expectations only to the extent that execution supports them. A conventional self-assessment signal uses oo alone and cannot distinguish an anticipated outcome from an unanticipated one. In Appendix A.3, the mean assessed outcome is 0.80 while SoA is 0.60, because the agent fulfilled one criterion it had not expected to. The reflection step acts on this discrepancy.

(2) Why the optimum p* lies above o. As stated in the introduction, agents are encouraged to expect better performance and revise their behavior to match those expectations. The optimum p∗p^* is the midpoint between the current outcome and full achievement, so the score favors a target above current performance by a bounded margin (1−o)/2(1 - o)/2 that vanishes as oo approaches 1. Predictions far above the outcome are penalized by the agreement term. A calibration-only objective with its optimum at p=op = o would provide no such direction. This choice is consistent with findings on human learning: Chambon et al. [1] report a positive learning asymmetry specific to self-determined choices and show in simulation that it improves performance.

(3) Whether the components behave as intended. If the margin acts as a target, the assessed outcome should rise toward the prediction, and the gap should narrow. Figure 3 shows this pattern for the verifier agent on BrowseComp-Plus: predictions and outcomes both increase, and the signed gap is generally smaller than at the start. We further report the per-round signed gap pˉ−oˉ\bar{p} - \bar{o} against the predicted optimal value (1−oˉ)/2(1 - \bar{o})/2, averaged over all agents and queries:

Table 3.1: Per-round signed gap pˉ−oˉ\bar{p} - \bar{o} and the predicted optimal value (1−oˉ)/2(1 - \bar{o})/2 on BrowseComp-Plus.

Roundpˉ\bar{p}oˉ\bar{o}pˉ−oˉ\bar{p} - \bar{o}(1−oˉ)/2(1 - \bar{o})/2
00.843 ± 0.0120.736 ± 0.0100.107 ± 0.0040.132 ± 0.005
10.897 ± 0.0140.757 ± 0.0140.140 ± 0.0050.122 ± 0.007
20.887 ± 0.0080.757 ± 0.0070.131 ± 0.0050.122 ± 0.004
30.884 ± 0.0010.760 ± 0.0100.123 ± 0.0090.120 ± 0.005
40.885 ± 0.0060.764 ± 0.0050.120 ± 0.0060.118 ± 0.002
50.885 ± 0.0140.760 ± 0.0140.125 ± 0.0130.120 ± 0.007
60.890 ± 0.0070.762 ± 0.0220.128 ± 0.0250.119 ± 0.011
70.896 ± 0.0070.778 ± 0.0150.118 ± 0.0120.111 ± 0.008
80.897 ± 0.0090.774 ± 0.0180.123 ± 0.0110.113 ± 0.009
90.895 ± 0.0080.776 ± 0.0120.119 ± 0.0050.112 ± 0.006

Observations: From the results, we can observe that: (1) the gap pˉ−oˉ\bar{p} - \bar{o} quickly converges near the optimum and then fluctuates slightly above it, indicating that the agents evolve in the direction that increases the current-round SoA score; and (2) pp always stays somewhat above oo, which empirically supports that the agents use a higher pp to guide their capabilities and thereby improve the actual outcome oo. We will also add the results in the manuscripts.

[1] Chambon et al. Information about action outcomes differentially affects learning from self-determined versus imposed choices. Nature Human Behaviour

AuthorsOpenReview ↗

Response to Weakness 2 and Question 1

Weakness 2: Although the paper includes SelfOrg (Tastan et al., ICLR 2026) and MetaGen (Wang et al., 2026) as internal-feedback baselines, these approaches involve different MAS evolution mechanisms and do not directly isolate the advantages of the proposed SoA signal over other forms of internal feedback. Existing approaches such as Reflexion and Self-Refine also demonstrate the effectiveness of self-assessment and self-reflection for iterative improvement. I would encourage the authors to consider additional controlled ablations within the same MAS evolution framework, replacing SoA with simpler internal feedback signals, such as (i) post-execution self-assessment alone, (ii) prediction–outcome discrepancy alone, or (iii) trajectory-based self-reflection without SoA. Such experiments would help establish whether the proposed SoA formulation provides additional benefits beyond standard internal feedback mechanisms.

Response: Thank you for your valuable comments. Following your suggestions, we further compare SOAP with Reflexion [2], Self-Refine [3], and COPPER [4], representative methods that use self-assessment and self-reflection for iterative improvement. In addition, we conduct further ablations with different feedback signals, as suggested in Weakness 2 and Question 1, to justify the effectiveness of the SoA feedback used in our method.

Table 3.2: Comparison with self-assessment and self-reflection methods. Results are reported as mean ± SD over three runs; on CloudCast-Haiku, n=10 denotes 10 trials.

MethodBrowseComp-PlusWorkBenchPlanCraftCloudCast-Haiku (n=10)
SOAP0.630 ± 0.0450.671 ± 0.0300.797 ± 0.0690.110 ± 0.011
Reflexion0.533 ± 0.0150.537 ± 0.0250.767 ± 0.0470.052 ± 0.014
Self-Refine0.543 ± 0.0510.568 ± 0.0210.770 ± 0.0360.095 ± 0.084
COPPER0.563 ± 0.0930.612 ± 0.0240.777 ± 0.0760.078 ± 0.061

Observations: From the results, we can observe that SOAP outperforms these three representative methods based on self-assessment and self-reflection. This further strengthens the empirical evidence for the superiority of our proposed method, and validates the effectiveness of extending the concept of "sense of agency" to encourage the agent to better understand what it can do and how its behavior will affect the outcome during evolution.

Table 3.3: Ablation on the feedback signal used for evolution. Results are reported as mean ± SD over three runs; on CloudCast-Haiku, n=10 denotes 10 trials.

Feedback SignalBrowseComp-PlusWorkBenchPlanCraftCloudCast-Haiku (n=10)
Outcome only (oo)0.577 ± 0.0320.563 ± 0.0070.750 ± 0.0360.092 ± 0.106
Prediction only (pp)0.597 ± 0.0580.548 ± 0.0120.723 ± 0.0320.048 ± 0.022
Agreement only (∣p−o∣\lvert p-o\rvert)0.613 ± 0.0400.611 ± 0.0340.750 ± 0.0890.053 ± 0.007
Trajectory0.537 ± 0.0320.583 ± 0.0120.753 ± 0.0210.063 ± 0.017
SOAP0.630 ± 0.0450.671 ± 0.0300.797 ± 0.0690.110 ± 0.011

Observations: From the results, we can observe that, compared with post-execution self-assessment alone, prediction–outcome discrepancy alone (i.e., ∣p−o∣\lvert p-o\rvert), and trajectory-based self-reflection without SoA, our method still achieves the best performance on all four benchmarks. This further justifies the effectiveness of leveraging the concept of "sense of agency", which encourages the agent to evolve toward a stronger sense of how it can behave and how its behavior will affect the outcome.

[2] Shinn et al. "Reflexion: Language Agents with Verbal Reinforcement Learning." Advances in Neural Information Processing Systems 36 (NeurIPS), 2023.

[3] Madaan et al. "Self-Refine: Iterative Refinement with Self-Feedback." Advances in Neural Information Processing Systems 36 (NeurIPS), 2023.

[4] Bo et al. "Reflective Multi-Agent Collaboration based on Large Language Models." Advances in Neural Information Processing Systems 37 (NeurIPS), 2024.

Question 1: Could the authors clarify the respective contributions of these two mechanisms? In particular, would the method still achieve similar improvements if the agents performed self-reflection using the same execution trajectories and outcome assessments, but without the proposed SoA scores?

Response: Thank you for your valuable questions. As presented in our response to Weakness 2 (Table 3.3), we conduct additional ablation studies to validate the contribution of SoA scores. We will also revise the manuscript accordingly.

AuthorsOpenReview ↗

Response to Question 2

Question 2: Interpretation of interaction-level SoA. The interaction-level SoA compares a sender's prediction of a downstream agent's performance with the receiver's self-assessed outcome. However, this does not necessarily measure the sender's actual contribution to the receiver's performance. Could the authors clarify why this metric is an appropriate measure of interaction-level agency?

Response: Thank you for your valuable question. We agree that interaction-level SoA does not isolate the sender's causal contribution to the receiver's performance, and we state this in Appendix A.3. The receiver's outcome also depends on its other inputs and its own execution. Below, we clarify what the metric measures, how the sender's contribution is measured separately, and why we adopt this design.

(1) What interaction-level SoA measures. What the score captures is whether a sender can anticipate the downstream effect of its output. In agency research, the match between a predicted and an obtained consequence of one's action is a cue to agency, and the interaction-level score applies this cue to a consequence located in another agent. A sender with low interaction-level SoA either expects little of the receiver or misjudges how its output will be used, and both cases indicate a handoff worth revising.

(2) How the sender's contribution is measured. Contribution is measured by a separate quantity. Topology edits combine the accumulated interaction-level SoA with marginal information coverage, which measures how much a sender's output covers the receiver's unmet rubric dimensions beyond what other sources provide (Appendix A.5). An edge is removed only when both are low, and low SoA alone does not justify removal when the sender still supplies needed information. The two signals are therefore complementary: SoA indicates whether the interaction behaves as the sender expects, and marginal coverage indicates whether the sender adds information.

(3) Why we adopt this design. We chose this design over a counterfactual assessment with the sender's message removed, which would require an additional execution per edge in every round. The ablation in Figure 4 shows that removing interaction-level SoA lowers performance on all four benchmarks. We will move the statement on causal contribution and the role of marginal coverage from the appendix into Section 3.4.2.

Reviewer_J2HZOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposed SOAP, an evolution framework for LLM-based multi-agent systems that uses the agent’s self-evaluation, rather than external validation or judgment, as the signal to drive adaptation. The authors borrowed the idea of “Sense of Agency” from cognitive science. Each agent creates rubrics per query, predicts its own and downstream agents’ achievements, then calculates individual and interaction SoA after execution. These signals then drive two evolutionary mechanisms: prompt word refinement through reflection, and rule-based edge addition and deletion in the collaboration graph. Experiments cover BrowseComp-Plus, WorkBench, PlanCraft, and CloudCast using Gemini-2.5-Flash as the default backbone.

Critical strengths

It’s timely and well-motivated. Avoiding reliance on external validation or judgment is a real obstacle to deploying self-improved MAS, and the introduction clearly explains this.

A clear two-level decomposition. Separating individual-level agent sense from interaction-level agent sense is a reasonable abstraction and naturally corresponds to two adaptation aspects.

The assessment has wide coverage. The four benchmarks span search, office tool usage, planning, and program optimization, along with several backbone models and team size scans.

Critical weaknesses

The objective function of SoA has a systematic optimistic optimum. The paper fails to rule out the possibility that the reported gains come from prediction/evaluation inflation rather than the true improvement of the agency. This is my core concern. A fixed outcome oo, maximizing A(p,o)=p(1−∣p−o∣)A(p,o) = p(1-|p-o|) over pp yields p∗=(1+o)/2p^* = (1+o)/2, i.e., the optimal prediction is always above the true outcome by (1−o)/2(1-o)/2, with optimal value (1+o)2/4(1+o)^2/4.

The most significant deficiency in the ablation experiment: feedback based solely on self-assessment. The argument of the paper is that the SoA, especially the prediction–outcome agreement term, drives improvement. However, the ablation seems to remove the entire feedback hierarchy rather than isolate the prediction items.

The main result includes no variance, random seed, or significance test. Table 1 reports only a single number for each cell, with no error bars, no nn, and no significance tests. Several claimed winning margins were so small that they might fall within the range of inter-run noise.

Questions for authors

A self-assessed control only. What exactly has been removed with "w/o Indiv.SoA "in Figure 4? If it removes the entire individual-level feedback, can a condition be added that retains the rubric-based self-assessment oo as the sole feedback signal (no pp, no gap item)?

arg⁡max⁡pA(p,o)=(1+o)/2>o\arg\max p A(p,o) = (1+o)/2 > o holds for all o<1o<1. Has your evolved agent converged to this optimistic optimum? Please report the comparison of the signed gap pˉ−oˉ\bar p -\bar o for each round with the predicted optimal value (1−oˉ)/2(1-\bar o)/2.

Why did SOAP, EvoAgent, and SelfOrg all score exactly 618.0? Is this the cost of the seed program? What are the values of nn, the standard deviation, and the proportion of trials that have improved compared to the seed?

6 comments
AuthorsOpenReview ↗

Response to Weakness 1 (part 1)

Thank you for your constructive comments, which help us improve our work. We have run additional experiments and provide further clarification below to address your concerns. Please let us know if we have misunderstood any of your comments, and we will respond promptly.


Weakness 1: The objective function of SoA has a systematic optimistic optimum. The paper fails to rule out the possibility that the reported gains come from prediction/evaluation inflation rather than the true improvement of the agency. This is my core concern. A fixed outcome o, maximizing A(p,o)=p(1−∣p−o∣)A(p, o) = p(1 - \lvert p - o\rvert) over p yields p∗=(1+o)/2p^* = (1 + o)/2, i.e., the optimal prediction is always above the true outcome by (1−o)/2(1 - o)/2, with optimal value (1+o)2/4(1 + o)^2/4.

Response: Thank you for your valuable comments. The derivation is correct: for a fixed oo, the score is maximized at p∗=(1+o)/2p^* = (1 + o)/2, and we will add this to the paper. An optimistic optimum would produce spurious gains only if two conditions held: the reported gains would have to be measured through pp or oo, and the evolution would have to move pp toward p∗p^*. We address each below.

(1) Rationality of the SoA score design and the systematic optimal prediction. As stated in the introduction, agents are encouraged to expect better performance and to revise their behavior so that execution matches those expectations. A prediction slightly above the current outcome creates a gap that the next round must close by improving execution, so the outcome is pulled toward a higher level.

  • Intuitively, the margin is bounded. p∗p^* is the midpoint between the current outcome and full achievement, the margin (1−o)/2(1 - o)/2 vanishes as oo approaches 1, and predictions far above the outcome are penalized by the agreement term. An objective with its optimum at p=op = o would reward agreement at any level, including low expectation with low achievement, and would give evolution no direction.
  • Theoretically, this design is consistent with findings on human learning. Chambon et al. [1] report a positive learning asymmetry for self-determined choices and show in simulation that it improves performance, and goal-setting research finds that goals set above current performance lead to higher achievement [2].

(2) Inflating p or o cannot produce the reported gains. Table 1 reports scores from an external evaluator on BrowseComp-Plus, WorkBench, and PlanCraft, and from a programmatic simulator on CloudCast. These measurements are independent of the agents' predictions and self-assessments, and on the first three benchmarks they are never supplied to SOAP (Appendix B.1). The improvements over the baselines are therefore improvements in task performance.

AuthorsOpenReview ↗

Response to Weakness 1 (part 2)

(3) The self-assessed outcome reflects real progress. We group queries by the external evaluator's judgment and compare the agents' final-round self-assessments on BrowseComp-Plus:

Table 2.1: Final-round self-assessments on BrowseComp-Plus, grouped by the external evaluator's judgment.

BenchmarkExternal judgment# QueriesMean ooMean ppMean SoA
BrowseComp-PlusCorrect / Incorrect73 / 770.824 / 0.7040.925 / 0.8480.801 / 0.680

Self-assessed outcomes are clearly higher on queries judged correct, so oo tracks external correctness. The derivation above holds oo fixed, whereas in SOAP oo changes across rounds. When the prediction exceeds the outcome, the reflection prompt (Template 4, Appendix A.2) asks the agent to revise its execution and instructs it not to inflate predicted scores. Execution improves, oo rises toward pp, and the higher outcome in turn supports a higher expectation. Both quantities move upward together, which is what raises SoA.

(4) The agents get close to the optimum at first and then fluctuate above it, while p keeps guiding the actual outcome closer to it. As requested in Question 2, we report, for each round, the signed gap pˉ−oˉ\bar{p} - \bar{o} against the predicted optimal value (1−oˉ)/2(1 - \bar{o})/2, averaged over all agents and queries on BrowseComp-Plus:

Table 2.2: Per-round signed gap pˉ−oˉ\bar{p} - \bar{o} and the predicted optimal value (1−oˉ)/2(1 - \bar{o})/2 on BrowseComp-Plus.

Roundpˉ\bar{p}oˉ\bar{o}pˉ−oˉ\bar{p} - \bar{o}(1−oˉ)/2(1 - \bar{o})/2
00.843 ± 0.0120.736 ± 0.0100.107 ± 0.0040.132 ± 0.005
10.897 ± 0.0140.757 ± 0.0140.140 ± 0.0050.122 ± 0.007
20.887 ± 0.0080.757 ± 0.0070.131 ± 0.0050.122 ± 0.004
30.884 ± 0.0010.760 ± 0.0100.123 ± 0.0090.120 ± 0.005
40.885 ± 0.0060.764 ± 0.0050.120 ± 0.0060.118 ± 0.002
50.885 ± 0.0140.760 ± 0.0140.125 ± 0.0130.120 ± 0.007
60.890 ± 0.0070.762 ± 0.0220.128 ± 0.0250.119 ± 0.011
70.896 ± 0.0070.778 ± 0.0150.118 ± 0.0120.111 ± 0.008
80.897 ± 0.0090.774 ± 0.0180.123 ± 0.0110.113 ± 0.009
90.895 ± 0.0080.776 ± 0.0120.119 ± 0.0050.112 ± 0.006

Observations: From the results, we can observe that: (1) the gap pˉ−oˉ\bar{p} - \bar{o} quickly converges to the vicinity of the optimum and then keeps fluctuating slightly above it, indicating that the agents do evolve in the direction that increases the SoA score of the current round; and (2) under this condition, pp always stays somewhat above oo, which empirically supports that the agents indeed use a higher pp to guide their capabilities, so that the actual outcome oo improves.

We will add these clarifications, explanations, and analyses to the revision.

[1] Valérian Chambon, Héloïse Théro, Marie Vidal, Henri Vandendriessche, Patrick Haggard, and Stefano Palminteri. "Information about action outcomes differentially affects learning from self-determined versus imposed choices." Nature Human Behaviour, 4(10): 1067–1079, 2020.

[2] Edwin A. Locke and Gary P. Latham. "Building a practically useful theory of goal setting and task motivation: A 35-year odyssey." American Psychologist, 57(9): 705–717, 2002.

AuthorsOpenReview ↗

Response to Weakness 2 and Question 1-3

Weakness 2 & Question 1: The most significant deficiency in the ablation experiment: feedback based solely on self-assessment. The argument of the paper is that the SoA, especially the prediction–outcome agreement term, drives improvement. However, the ablation seems to remove the entire feedback hierarchy rather than isolate the prediction items. What exactly has been removed with "w/o Indiv.SoA" in Figure 4? If it removes the entire individual-level feedback, can a condition be added that retains the rubric-based self-assessment o as the sole feedback signal (no p, no gap item)?

Response: Thank you for your valuable comments and constructive questions. Following your suggestion, we add new experiments with self-assessment-only feedback. The settings are as follows: (1) the ablation in the manuscript (w/o Indiv. SoA) removes the individual-level SoA score and keeps only the interaction-level SoA score; and (2) the new condition (Self-assessment alone) replaces the individual-level SoA with self-assessment under the same rubrics used for Indiv. SoA. The results are as follows:

Table 2.3: Ablation on individual-level feedback. Results are reported as mean ± SD over three runs; on CloudCast-Haiku, n=10 denotes 10 trials.

MethodBrowseComp-PlusWorkBenchPlanCraftCloudCast-Haiku (n=10)
SOAP0.630 ± 0.0450.671 ± 0.0300.797 ± 0.0690.110 ± 0.011
w/o Indiv. SoA0.570 ± 0.0360.563 ± 0.0450.733 ± 0.0120.063 ± 0.073
Self-assessment alone0.577 ± 0.0320.563 ± 0.0070.750 ± 0.0360.092 ± 0.106

Observations: From the results, we can observe that: (1) compared with self-assessment alone, using the SoA score proposed in our method as feedback brings larger performance gains, which validates that reflecting on both the pre-execution prediction and its gap from the actual performance (i.e., ∣p−o∣\lvert p-o\rvert) helps guide agent evolution; and (2) entirely removing individual-level feedback (w/o Indiv. SoA) performs worse than self-assessment feedback in most cases. This is reasonable, since self-assessment still provides feedback on the execution outcome, which encourages the agent to learn from failed executions and self-improve; and (3) on CloudCast-Haiku, a coding task whose correctness is relatively easy to verify, self-assessment alone is fairly competitive; on the other, more complex tasks, however, its advantage is less evident.


Question 2: arg⁡max⁡pA(p,o)=(1+o)/2>o\arg\max_p A(p, o) = (1 + o)/2 > o holds for all o<1o < 1. Has your evolved agent converged to this optimistic optimum? Please report the comparison of the signed gap pˉ−oˉ\bar{p} - \bar{o} for each round with the predicted optimal value (1−oˉ)/2(1 - \bar{o})/2.

Response: Thank you for your insightful question. Following your suggestion, we report the signed gap pˉ−oˉ\bar{p} - \bar{o} and the predicted optimal value (1−oˉ)/2(1 - \bar{o})/2 for each round in Table 2.2, and analyze the observations in our response to Weakness 1.


Question 3: Why did SOAP, EvoAgent, and SelfOrg all score exactly 618.0? Is this the cost of the seed program? What are the values of n, the standard deviation, and the proportion of trials that have improved compared to the seed?

Response: Thank you for your question. CloudCast is a program-optimization task, and the cost of the seed program is 1035 (Appendix B.2). With Claude Sonnet 5 as the execution backbone, 618.0 is the best cost observed in practice on this benchmark. The stronger baselines also reach this level, which we attribute to the strength of the backbone. To better distinguish the methods on this benchmark, we reran all of them with a weaker backbone, Claude Haiku, and update the main results as follows:

Table 2.5: Results on CloudCast with Claude Haiku. n is the number of trials; the score is computed as (1035 − cost)/1035. A higher score (i.e., a lower cost) denotes better performance. "Improved over seed" is the number of trials whose cost is lower than that of the seed program.

MethodnScore (mean ± SD)Average costImproved over seed
EvoAgent100.020 ± 0.01410146/10
SelfOrg100.045 ± 0.0249889/10
SOAP100.110 ± 0.011921.310/10

Observations: From the results, we can observe that SOAP outperforms these competitive baselines. SelfOrg achieves a lower cost than EvoAgent, mainly because SelfOrg focuses on coordinating different agent roles rather than generating new agents with relatively weak capabilities. SOAP outperforms both, which supports the effectiveness of our method.

Reviewer_PDNsOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes SOAP, a framework for inference-time adaptation of multi-agent systems. Agents generate fixed, task- and role-specific rubrics, predict their own and downstream agents' performance, and assess completed execution against the same rubrics. A computational sense-of-agency score, p(1−|p−o|), combines anticipated achievement with prediction–outcome agreement. Individual feedback guides prompt updates, while accumulated interaction feedback and marginal information coverage guide topology edits. The paper evaluates SOAP on information retrieval, workplace tool use, crafting, and program optimization. Table 1 reports the highest accuracy among the compared methods on the first three benchmarks and a tie for the lowest CloudCast cost.

Critical strengths

The presentation is good for this paper as I followed through it pretty well.

The paper addresses a relevant problem, enabling multi-agent systems to adapt to individual tasks while reducing reliance on separately designed evaluation mechanisms.

The reported results across multiple tasks support the usefulness of the overall framework, while the ablations provide evidence for the contributions of individual-level and interaction-level adaptation.

Critical weaknesses

The paper addresses a practically relevant engineering problem: how feedback generated by agents themselves can guide task-specific changes to instructions and collaboration structures, thereby reducing dependence on separately constructed evaluation mechanisms. It combines prospective performance assessment, retrospective self-evaluation, and updates at both the individual and interaction levels. The empirical results support the usefulness of this overall framework. My main concern is whether the connection between this engineering objective and the proposed sense-of-agency interpretation is sufficiently established. The results show that the framework is useful, but they do not make equally clear why its effectiveness should be understood in terms of agency, or what this concept explains beyond more general mechanisms of self-evaluation and reflection.

First, I would appreciate a more explicit statement of the central research question (the "WHY"/"WHAT" question, not the "HOW" question). The introduction asks: “A central question is therefore what feedback should guide this evolution.” This is a valid question, but it does not by itself establish sense of agency as the appropriate feedback mechanism. Several steps separate the need for internal feedback, the potential usefulness of combining prospective and retrospective assessments, and the interpretation of a particular combination of these assessments as agency.

The motivation for the first step is reasonably clear, whereas the subsequent steps rely more heavily on intuitive analogies. What hypothesis is the paper intended to test?? That paired prospective and retrospective assessments outperform ordinary self-reflection? That the particular scoring function improves adaptation? Or that an agency mechanism explains how multi-agent systems adapt? These are related but distinct questions!!!

Second, the relationship to metacognition, self-monitoring, and uncertainty-related research needs clearer positioning. Functionally, SOAP predicts its own performance, examines execution outcomes, compares the two assessments, and uses the resulting information to modify subsequent behavior. This has a clear connection to the distinction between monitoring and control in metacognitive frameworks. Language-model research has also examined both judgments about the correctness of a particular answer and predictions about whether a model can answer a question. Of course, SOAP defines (p) as anticipated achievement rather than an explicit probability of correctness, so it should not simply be equated with an uncertainty estimator. The more relevant question is what the agency interpretation adds to these existing mechanisms of prospective judgment, retrospective monitoring, and strategic adjustment.

Regarding the cognitive-science foundation, my concern is not that the authors have failed to cite agency research. The paper references relevant work by Haggard, Moore and Fletcher, Chambon, and others. Rather, the correspondence between the mechanisms discussed in that literature and the variables used here is insufficiently developed. Comparator or forward-model accounts concern the specific consequences of an action and the relationship between predicted consequences and actual feedback; this is not automatically equivalent to agreement between two assessments of task achievement. Cue-integration accounts additionally consider the reliability of different cues and prior information, while formal Bayesian models of agency can explicitly distinguish action causation from alternative causal explanations. SOAP need not implement all these details, but it should explain which mechanism its simplification retains and why that mechanism supports the proposed interpretation.

In particular, it is necessary to distinguish “I expect to perform well,” “I accurately anticipated the outcome,” and “I experience this outcome as caused or controlled by my action.” These judgments may be related, but they are not interchangeable. Cognitive experiments also show that perceived control can be influenced by the fluency of action selection and can dissociate from execution performance. Combining high anticipated achievement with prediction–assessment agreement may therefore be a reasonable planning heuristic, but further justification is needed before treating it as a valid proxy for agency.

These concerns lead to two focused requests. First, clarify the research question and the intended contribution, including the relationship to metacognition, self-monitoring, and reflective methods. If the main contribution is an agency-inspired engineering approach, that would be a reasonable positioning, provided that the theoretical interpretation is aligned with the available evidence. This does not require a complete model of human cognition or a demonstration that the proposed scoring function is uniquely justified. It does require a clear account of what role the agency concept plays in the design and which claims about agency remain outside the scope of the study.

My reservations primarily concern the theoretical interpretation and attribution of the contribution. The reported results make the internal-feedback and adaptation framework worthy of serious consideration. At the same time, a gap remains between showing that the system improves and showing that agency explains that improvement. How clearly the authors address this gap would materially affect my assessment of the paper’s originality and the understanding it adds beyond existing approaches to self-evaluation and reflection!!

【 A further change of rating is likely, if authors successfully address my concerns ! ! ! 】

Questions for authors
  1. Research question and conceptual positioning.

What specific hypothesis motivates interpreting the proposed feedback as sense of agency, beyond prospective performance assessment and retrospective self-reflection? Although the paper cites relevant agency research, the connection between these theories and the proposed score remains unclear. Could you clarify which aspect of agency the score captures, how it differs from metacognitive monitoring and existing self-refinement approaches, and whether the intended contribution is an agency-inspired engineering heuristic or a stronger computational account of agency?

7 comments
AuthorsOpenReview ↗

Response to Weaknesses and Questions (part 1)

Dear Reviewer PDNs, thank you for your careful and constructive reading. Your review separates what our results establish from what our framing suggests, and this distinction has helped us sharpen the paper. We understand your main concerns to be twofold: (1) the motivation for using the sense of agency, including how SOAP maps onto agency theory and what benefits SoA brings as a feedback signal; and (2) how SOAP differs from existing self-assessment and metacognitive monitoring methods. To answer your final question directly: we position SOAP as an agency-inspired engineering approach, in which research on the sense of agency (SoA) guides the design of the internal feedback signal. The response below follows your line of reasoning: the research question and contribution, why SoA is a suitable source of internal feedback, which mechanisms the design retains, how it relates to metacognition and reflective methods, and what the evidence supports.


1. Research question and intended contribution

The research question is: when no external evaluator or verification environment is available, what signal can agents produce by themselves that is informative enough to guide the test-time evolution of both individual behavior and collaboration structure?

We aim to justify two hypotheses:

  • Hypothesis 1: Feedback that pairs a pre-execution prediction with a post-execution assessment under the same rubric guides evolution better than retrospective self-assessment alone.
  • Hypothesis 2: The proposed SoA score, i.e., p(1−∣p−o∣)p(1-\lvert p-o\rvert), applied at both the individual and interaction levels, is an effective objective for this adaptation.

The agency concept plays a design role in SOAP. It determines which quantities we elicit (a prediction before acting and an assessment after acting), how we combine them (expectation weighted by agreement), and where we apply them (to an agent's own execution and to the effect of its output on downstream agents). We agree that the introduction moves from the need for internal feedback to SoA mainly by analogy, and we will make these steps explicit in the revision. As for whether an agency mechanism accounts for how multi-agent systems adapt, we believe it is worth exploring; it is a deeper question that we regard as a direction for future work.


2. Why SoA is a suitable source of internal feedback

Our engineering objective imposes two requirements on a feedback signal: (1) it must be available without external rewards or judgments, and (2) it must indicate what to change, i.e., which agent or which connection is responsible for an outcome. Research on human agency reports signals that serve both functions, which is why we turn to SoA.

  • (1) Regarding the first requirement, evidence of control over an effect reinforces action in the absence of external reward. Eitam et al. [1] show that an action's own effect increases motivation, and that this increase is not explained by additional information about task performance. Karsh & Eitam [2] and Karsh et al. [3] show that agency judgments and control-relevant information influence subsequent action selection. A point-by-point mapping of component in our proposed SoA score and the concept in cognitive science is presented in the next section (Table 1.1).

  • (2) Regarding the second requirement, agency shapes how people learn from outcomes. Chambon et al. [4] find that a learning-rate asymmetry appears for self-determined choices but not for imposed ones, and that this asymmetry improves performance in simulation. Di Costa et al. [5] and Majchrowicz et al. [6] report that an implicit measure of agency increases after errors only when the errors can be used to improve later choices. In motor learning, Berniker & Körding [7] model adaptation as estimating the source of an error, i.e., a credit assignment problem between one's own body and the world. Since credit assignment across agents and edges is the central difficulty in MAS evolution, this line of work is the closest conceptual counterpart to our setting.

Nonetheless, these studies mostly use implicit measures or low-level action effects, whereas SOAP uses explicit judgments. Moreover, the relation between agency and learning is not monotonic: Seghezzi et al. [8] find that intentional binding decreases over the course of learning. We therefore draw on this literature as motivation for our design, and we regard establishing the score as a measure of agency as a separate question.


AuthorsOpenReview ↗

Response to Weaknesses and Questions (part 1 references + part 2)

3. Which mechanisms the design retains

Table 1.1 summarizes the correspondence we intend and where SOAP simplifies the theory.

Table 1.1: Correspondence between SOAP components and their counterparts in agency research.

SOAP componentCounterpart in agency researchWhat is retainedWhere SOAP simplifies
ppForward-model prediction of an action's consequenceAn expectation formed before actingPredicts rubric-level achievement instead of sensory consequences
ooActual feedbackAn outcome signal obtained after actingSelf-assessed instead of external feedback
1−∣p−o∣1-\lvert p-o\rvertComparator congruenceMatch between prediction and outcome as a cue to agencyA judgment-level comparison of two assessments
pi→jp_{i\to j} with ojo_jEffect of an action on the external worldThe consequence lies outside the agent and is reported by another agentThe receiver's report is also a self-assessment
SoA-guided updatesAgency judgments influencing action selectionMonitoring output is used to control later behaviorRealized through prompt and edge revision

SOAP retains one mechanism from comparator accounts [9, 10]: congruence between a predicted and an obtained outcome as a cue to agency. This mechanism operates at the level of explicit judgment, in the sense of judgment of agency as distinct from feeling of agency [11]. Reliability-weighted cue integration and Bayesian inference over alternative causes are richer mechanisms that our simplification leaves for future extensions.

Regarding the three judgments you distinguish, we agree that they are not interchangeable. "I expect to perform well" corresponds to pp, which, taken alone, is closer to an outcome expectation. "I accurately anticipated the outcome" corresponds to 1−∣p−o∣1-\lvert p-o\rvert, the comparator-style cue. "I experience this outcome as caused or controlled by my action" is the judgment that our individual-level score approximates least. The interaction-level score comes closest to it, because the predicted consequence is located in another agent and reported by that agent; we therefore treat it as a proxy. Likewise, the influence of action-selection fluency on perceived control [12] is beyond what the current score models.

The pp term serves an engineering purpose. Agreement alone admits a degenerate solution in which an agent predicts low and performs low, thereby reaching perfect agreement. Weighting by pp assigns this case a low score (zero when p=0p=0) and directs evolution toward higher achievement that the agent can also anticipate.

[1] Eitam et al. Motivation from control. Experimental Brain Research, 2013.

[2] Karsh et al. I control therefore I do: Judgments of agency influence action selection. Cognition, 2015.

[3] Karsh et al. Bootstrapping agency: How control-relevant information affects motivation. Journal of Experimental Psychology: General, 2016.

[4] Chambon et al. Information about action outcomes differentially affects learning from self-determined versus imposed choices. Nature Human Behaviour, 2020.

[5] Di Costa et al. Try and try again: Post-error boost of an implicit measure of agency. Quarterly Journal of Experimental Psychology, 2018.

[6] Majchrowicz et al. Learning from informative losses boosts the sense of agency. Quarterly Journal of Experimental Psychology, 2020.

[7] Berniker et al. Estimating the sources of motor errors for adaptation and generalization. Nature Neuroscience, 2008.

[8] Seghezzi et al. Intentional binding decreases during learning: Implications for sense of agency. Quarterly Journal of Experimental Psychology, 2026 (published online 2025).

[9] Chambon et al. From action intentions to action effects: How does the sense of agency come about? Frontiers in Human Neuroscience, 2014.

[10] Moore et al. Sense of agency in health and disease: A review of cue integration approaches. Consciousness and Cognition, 2012.

[11] Synofzik et al. Beyond the comparator model: A multifactorial two-step account of agency. Consciousness and Cognition, 2008.

[12] Chambon et al. Sense of control depends on fluency of action selection, not motor performance. Cognition, 2012.


AuthorsOpenReview ↗

Reponse to Weaknesses and Questions (part 3)

4. Relation to metacognition, self-monitoring, uncertainty, and reflective methods

Individual-level SoA is an instance of the monitoring-and-control loop described in metacognitive frameworks [13]: the agent monitors its performance and uses the result to control its later behavior. We will state this connection in the paper.

  • What the agency framing adds is design structure. The object of monitoring shifts from the agent's knowledge state to the consequence of its action. This shift leads to the interaction-level score, where a sender predicts how a downstream agent will perform after receiving its output. Metacognitive monitoring of one's own answers has no direct counterpart to this quantity, and it is this signal that drives topology edits. The framing also couples the level of expectation with agreement in a single objective, which gives the update a direction: when pp exceeds oo, the agent revises its strategy; when oo exceeds pp, it raises its expectation to match its demonstrated capability.

  • Regarding uncertainty research, confidence in an answer and predictions of answerability are judgments about a single response or about knowledge. As you note, pp is anticipated achievement rather than a probability of correctness. It is elicited per rubric dimension before execution, paired with a post-execution assessment on the same dimension, and used as an optimization target for evolution.

  • Regarding reflective methods such as Reflexion [14], these rely on retrospective assessment. SOAP adds a prospective prediction committed before execution, so the feedback is the discrepancy between two assessments rather than a single post hoc critique. We have also included self-reflection methods (Reflexion [14] and Self-Refine [15]), a self-assessment method (COPPER [16]), and a metacognitive monitoring method as baselines for comparison. Metacognitive Monitoring is a baseline we implemented following the monitoring–control structure of metacognition [13]: after execution, each agent reports its confidence in its own output, and low confidence triggers a revision of its configuration, with no pre-execution prediction and no rubric. The results are as follows:

Table 1.2: Comparison of SOAP with self-reflection, self-assessment, and metacognitive monitoring baselines. "(MAS)" denotes that we adapt the method to the multi-agent system, where each agent performs self-reflection. Results are reported as mean ± SD over three runs; on CloudCast-Haiku, n=10 denotes 10 trials.

MethodBrowseComp-PlusWorkBenchPlanCraftCloudCast-Haiku (n=10)
SOAP0.630 ± 0.0450.671 ± 0.0300.797 ± 0.0690.110 ± 0.011
Reflexion (MAS)0.533 ± 0.0150.537 ± 0.0250.767 ± 0.0470.052 ± 0.014
Self-Refine (MAS)0.543 ± 0.0510.568 ± 0.0210.770 ± 0.0360.095 ± 0.084
COPPER0.563 ± 0.0930.612 ± 0.0240.777 ± 0.0760.078 ± 0.061
Metacognitive Monitoring0.503 ± 0.0120.587 ± 0.0010.733 ± 0.0310.053 ± 0.052

Observations: From the results, we can observe that SOAP outperforms the self-reflection, self-assessment, and metacognitive monitoring methods, which empirically validates the effectiveness of leveraging the sense of agency for self-evolution. On CloudCast-Haiku, a coding task whose correctness is relatively easy to verify, the self-reflection and self-assessment baselines are fairly competitive; on the other, more complex tasks, however, their advantage is less evident.

[13] Nelson et al. Metamemory: A theoretical framework and new findings. Psychology of Learning and Motivation (Academic Press), 1990.

[14] Shinn et al. Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems 36 (NeurIPS), 2023.

[15] Madaan et al. Self-Refine: Iterative Refinement with Self-Feedback. Advances in Neural Information Processing Systems 36 (NeurIPS), 2023.

[16] Bo et al. Reflective Multi-Agent Collaboration based on Large Language Models. Advances in Neural Information Processing Systems 37 (NeurIPS), 2024.


Comments on the paper

1 comment
AuthorsOpenReview ↗

Author Summary for the Area Chair

Dear Area Chair,

Thank you for handling our submission. We summarize our contributions, how the main concerns were addressed, and the discussion status.

1. Contributions

  • Internal feedback for multi-agent evolution. Inspired by the human sense of agency, each agent predicts its achievement before acting, assesses it afterward under the same rubric, and uses expectation weighted by prediction-outcome agreement as feedback, with no external verifier or LLM evaluator. Reviewers found the problem relevant (PDNs), timely and well-motivated (J2HZ), and the idea interesting (hRFt).
  • Two-level feedback. The signal guides prompt updates (individual level) and topology edits (interaction level); ablations show both contribute on all four benchmarks.
  • Broad evaluation. On four diverse benchmarks, SOAP is effective, outperforming a wide range of baselines, including methods using external feedback; robust across LLM backbones and team sizes; and interpretable, since its predictions, outcomes, and SoA scores show how each agent evolves. Analysis of SoA dynamics during evolution shows that individual- and interaction-level SoA rise together with task performance, and that predictions and assessed outcomes rise while their gap narrows. These results provide empirical support for the agency-inspired concepts behind SOAP.

2. How the Main Concerns Were Addressed

ConcernRaised byStatus
(1) Motivation and mapping to cognitive sciencePDNsResolved; rating raised
(2) Comparison with self-assessment/reflectionPDNs, J2HZ, hRFtNew experiments
(3) SoA score favors predictions above outcomesJ2HZ, hRFtClarified; new evidence
(4) Run-to-run noiseJ2HZThree seeds

(1) Motivation for sense of agency and mapping to cognitive science (PDNs). Resolved. We position SOAP as an agency-inspired engineering approach, state two explicit hypotheses, and map each score component to agency research (pp to forward-model prediction, 1−∣p−o∣1 - |p - o| to comparator congruence), noting where we simplify. PDNs raised their rating to weak accept.

(2) Comparison with self-assessment and self-reflection (PDNs, J2HZ, hRFt). New experiments.

  • SOAP outperforms Reflexion, Self-Refine, COPPER, and a metacognitive-monitoring baseline on all four benchmarks.
  • With the evolution mechanism fixed, SOAP beats outcome-only, prediction-only, agreement-only, and trajectory-based reflection feedback on all four benchmarks.
  • PDNs considered this resolved. We ran every follow-up variant hRFt requested, including removing only the computed score.

(3) For a fixed outcome o, the SoA score is maximized at a prediction above it, p* = (1 + o)/2 > o, which may favor optimistic predictions over calibrated ones and inflate the reported gains (J2HZ, hRFt). Clarified with new evidence. This is intentional: a slightly higher prediction acts as a target that guides evolution, consistent with cognitive-science findings that self-determined learning benefits from a positive asymmetry and that goals above current performance raise achievement. The margin is bounded and vanishes as o→1o \to 1. An optimum at p=op = o would reward low expectation with low achievement and give no direction. Gains do not come from inflation:

  • All reported results come from an external evaluator or simulator that SOAP never sees.
  • Self-assessed outcomes track external correctness.

(4) Run-to-run noise (J2HZ). New experiments.

  • All methods rerun with three random seeds. SOAP, using no external feedback, ranks first on BrowseComp-Plus, WorkBench, and CloudCast, and is on par with MetaAgent on PlanCraft, which uses external feedback in training.
  • With a weaker backbone on CloudCast, SOAP improves on the seed program in every trial, more consistently than the baselines.

We will incorporate all clarifications and new results into the paper.

3. Discussion Status

  • PDNs raised the rating to weak accept.
  • hRFt engaged actively, and we provided every ablation they requested, which we believe addresses their main concern, but the reviewer did not indicate whether concerns remain or whether they would revise their rating.
  • J2HZ, who raised concerns (2) to (4), has not responded despite two follow-ups.

For both weak-reject reviews, it is unclear which concerns, if any, remain.

4. Concluding Remarks

The new evidence confirms that SOAP's gains stem from its agency-inspired feedback, offering a practical path for multi-agent systems to evolve on their own without external verifiers. We hope the strengthened evidence and clarifications will be taken into account in your overall assessment of the paper.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.