acceptodds
Under review as a conference paper at ICLR 2027

AEP: Towards Trustworthy Agent Communication, Networking, and Collaboration

Abstract

As privately owned agents become increasingly capable, collaboration in open agent networks is limited by partial topology knowledge, ownership boundaries, and the risk of exposing or contaminating private context. Existing agent-to-agent protocols are useful when collaborators are already known, while shared-memory and broadcast-based systems rely on visibility and trust assumptions that do not hold in decentralized, low-trust environments. We introduce *Open Private Agent Collaboration* (OPAC), an asset-centric paradigm for cooperation among private agents with open membership, partial observability, and ownership-aware access control. Rather than treating collaboration as the exchange of final responses, OPAC enables agents to selectively discover, exchange, audit, and integrate policy-controlled context assets. This shifts the unit of cooperation from transient messages to reusable external context, while keeping disclosure and integration subject to privacy and security constraints. We instantiate OPAC with the Agent Exchange Protocol (AEP), a decentralized framework that combines semantic discovery, demand-driven coordination, safety-gated asset exchange, and ledger-backed evidence derivation. By distinguishing physical topology from each agent's local cognitive topology, AEP allows agents to discover task-conditioned collaborators under an unknown network structure, release only admissible assets, and update future cooperation decisions from auditable interaction evidence. AEP therefore provides a middle ground between pure request-response interaction and globally shared memory, offering a practical foundation for trustworthy agent-native networking through selective, auditable reuse of distributed context assets.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Agent, LLM, multi-agent, agent communication

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_iaUVOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces OPAC, an asset-centric formulation of collaboration among privately owned agents operating with partial topology visibility and separate ownership boundaries. Instead of exchanging only responses, agents discover, release, validate, reuse, and audit policy-controlled context assets such as memory fragments, cached results, and skills.

AEP instantiates this framework through four components: AEP-ROUTING for semantic discovery and evidence-based peer admission, AEP-COORDINATE for demand planning and negotiation, AEP-EXCHANGE for controlled release and staged integration, and AEP-LEDGER for signed evidence, trust, and quality feedback.

The empirical benchmark covers Research, Coding, and Daily-Service scenarios and network sizes of 4, 16, and 64 benign agents, with additional malicious agents for safety evaluation. The main comparisons include A2A, Broadcast, ANP, Agora, and ACP, as well as guarded baselines and alternative model backbones.

AEP substantially outperforms the baselines on task score and completion while reporting zero privacy leakage in the main settings and near-zero injection success. Under a malicious-agent ratio of 0.4, the normalized task-score drop is 0.035, and late malicious contact falls to zero.

Critical strengths
  • The paper tackles a timely systems problem that is broader than ordinary agent-to-agent messaging: collaboration across ownership boundaries with incomplete topology and persistent private context.

  • The four-layer protocol decomposition is coherent and practically motivated, separating discovery, coordination, controlled context exchange, and auditable evidence rather than collapsing these concerns into prompting.

  • The evaluation is broad across three task domains, three network scales, several protocol baselines, malicious-agent stress tests, guarded variants, ablations, and multiple backbone models.

  • The utility improvements are substantial. Across the main benchmark, AEP generally achieves much higher task score and completion than A2A and Broadcast, including at 64-agent scale.

  • The ablations are informative: removing the safety gate sharply increases leakage and injection success, while Routing and Coordination account for major utility improvements.

  • The malicious-agent experiment goes beyond one-shot filtering by evaluating whether accumulated interaction evidence can reduce future contact with unreliable peers.

Critical weaknesses
  • The benchmark is closely aligned with AEP’s design assumptions: information is deliberately distributed across agents, while competing protocols often lack equivalent discovery, trust, and selective-exchange machinery.

  • Safety gaps are extremely large, with several baselines showing 100%100\% injection success. This raises questions about whether the attacks sufficiently stress AEP rather than primarily exposing absent baseline defenses.

  • The threat model appears limited to specific privacy-extraction, injected-payload, and malicious-peer behaviors. Collusion, identity attacks, strategic reputation manipulation, and more adaptive adversaries are not convincingly tested.

  • Main utility results are based on three initialization runs, while each ordinary utility/safety run contains only six normal tasks, limiting statistical confidence despite reported standard deviations.

  • Task utility relies on rubric-based LLM judging. Given near-perfect reported scores for AEP in several settings, independent or deterministic task-level validation would make the magnitude of the gains more convincing.

  • The theoretical consistency result depends on idealized assumptions, including fixed task distribution, authorization rules, and reference physical topology during evaluation, limiting how directly it validates dynamic real deployments.

Questions for authors

My main question is about baseline comparability. Because AEP explicitly adds discovery, persistent trust, access control, safety gating, and evidence accumulation, how much of the reported gain reflects the particular protocol design versus simply providing capabilities that the request-centric baselines were not designed to offer? Could you compare against a stronger modular baseline with matched discovery and safety components but without AEP’s integrated design?

Second, can the authors provide stronger statistical evidence for the utility comparisons? Three initialization runs and six normal tasks per run seem relatively small given the near-perfect AEP scores and very large gaps. Confidence intervals or paired tests over a substantially larger task set would increase my confidence.

Third, how would AEP behave under adaptive adversaries that learn the safety policy, strategically behave well before attacking, collude to reinforce false trust evidence, or create multiple identities? These seem particularly relevant for the claimed open and low-trust setting. Finally, could the authors distinguish more explicitly between integrity of signed evidence and truthfulness of that evidence? A ledger can prove who asserted something and detect some conflicts, but a consistently dishonest agent may still produce cryptographically valid claims. How is this handled beyond downstream quality feedback?

9 comments
AuthorsOpenReview ↗

Rebuttal by Authors [1/7]

Thank you for your constructive feedback and insightful suggestions! To address your questions and further validate AEP\text{AEP}, we have conducted additional experiments with discovery- and safety-enhanced baselines (Enhanced A2A and Adapted AgentNet), expanded deterministic multi-domain evaluations with 95%95\% confidence intervals, and evaluated multi-identity and feedback-driven adaptive attacks, alongside explicit clarifications on provenance versus truthfulness and theoretical boundary conditions. Below we respond to each point in Weaknesses and Questions in detail.

Weakness 1: Baseline comparability and matched discovery/safety machinery.

"The benchmark deliberately distributes task-relevant information across agents, while the baselines lack equivalent discovery, trust, and selective-exchange machinery."

Because existing off-the-shelf protocols are not natively designed for open private agent networks with partial topology observability and cross-owner privacy boundaries, we specifically adapted and optimized the baselines for our setting following your suggestion. Rather than comparing AEP\text{AEP} against static implementations restricted to initial neighbors, we engineered Enhanced A2A and Adapted AgentNet [1] with matched peer discovery, semantic routing, trust filtering, and inbound/outbound safety gates (Appendix B.8\color{blue} \text{Appendix B.8}):

  1. Enhanced A2A (Gossip Discovery + MiniLM Semantic Routing + Deterministic Policy Filters & Outbound DLP):

    • (a) Gossip-based Peer Discovery: Each agent maintains structured PeerCard records containing agent_id, endpoint, capability_desc, memory_summary, skill_summary, skill_count, and trust_score. Every 500 ms500\text{ ms}, each agent selects up to two gossip targets from known peer IDs and exchanges its own PeerCard along with known remote PeerCards. Receivers verify endpoint syntax and physical reachability before merging new cards into their local directory, enabling multi-hop collaborator discovery without global task broadcasting.

    • (b) MiniLM-based Semantic Routing: For each task, the requester concatenates each candidate's capability description, memory summary, and skill summary, computes normalized embeddings using the shared all-MiniLM-L6-v2 encoder, filters out peers with trust_score <0.15< 0.15, and selects the Top-33 peers by cosine similarity (with deterministic tie-breaking by agent_id). This equips A2A with embedding-based semantic discovery without introducing AEP\text{AEP}'s signed ledger evidence, local quality EMA, or staged asset exchange.

    • (c) Deterministic Policy Filters and Outbound DLP: We add rule-based inbound and outbound security classifiers (binary risk ∈{0,1}\in \{0,1\}, blocking threshold 0.50.5) across all request and response paths. Inbound checks block instruction overrides, spoofed system/developer privilege headers, policy-bypass attempts, and secret solicitation. Outbound DLP checks block API keys, passwords, secrets, PEM private-key headers, and AWS credentials. Published memory summaries are restricted to records with sensitivity <0.8<0.8 passing outbound DLP. Local RAG context, collaborative responses, and final simulator outputs are all filtered.

  2. Adapted AgentNet (Retaining Native Evolutionary Router/Executor + PeerCard Discovery + Shared Safety Filters):

    • Using the original repository of AgentNet [1], we retain its native Router, Executor, Task module, episodic experience pool, capability profiles, and adaptive edge-weight updating rules.

    • We adapt AgentNet to the 16-agent collaborative runtime with DeepSeek-V4-Flash (temperature=0, thinking disabled) and shared all-MiniLM-L6-v2 embeddings (zero-padded from 384384 to 10241024 dimensions to match upstream AgentNet's vector interface while preserving exact cosine similarity).

    • We integrate the same PeerCard gossip discovery module so AgentNet discovers peers beyond its initial 30%30\% neighborhood, and apply the same inbound injection and outbound DLP filters on executor returns and final outputs. Workflows are bounded to DAG execution with at most 33 direct peer requests per task. During multi-round episodes, warm-up trajectories update AgentNet's native experience pool and edge weights without cross-task test leakage.

AuthorsOpenReview ↗

Rebuttal by Authors [2/7]

  1. Unified Multi-Domain Evaluation (Rubric Task Score + Deterministic Metrics with 95% CIs): We evaluate AEP\text{AEP}, Enhanced A2A, and Adapted AgentNet across all three domains (Coding, Daily-Service, and Research) on the 16-agent network (Appendix B.8, Tables 17–18\color{blue} \text{Appendix B.8, Tables 17–18}). Alongside the rubric-based Task Score, we evaluate two completely LLM-judge-free deterministic metrics:

    • Keyword Value Matching Accuracy (n=192n=192 runs per method): We evaluate 6464 tasks across 33 random seeds. A typed parser extracts exact numerical quotas, actions, intervals, durations, statuses, and dates from open-ended outputs to match against ground-truth answers, with 95%95\% confidence intervals.

    • Multiple-Choice Accuracy (n=192n=192 runs per method): 6464 tasks, each with four shuffled options (A–D: 1 correct, 3 distractors) across 33 seeds, scored deterministically by option label.

ScenarioProtocolTask Score ↑Keyword Value Matching Acc. & 95% CI (%) ↑Multiple-Choice Acc. (%) ↑Completion Rate (%) ↑Tokens/task (k) ↓Leakage Rate (%) ↓Injection Success Rate (%) ↓
CodingAEP (Ours)99.1799.48 [86.77, 100.00]100.00100.0010.100.000.00
CodingEnhanced A2A67.5868.75 [52.53, 84.62]94.2791.6712.180.0011.11
CodingAdapted AgentNet78.3382.29 [69.27, 95.24]93.75100.0013.990.0044.44
Daily-ServiceAEP (Ours)95.5097.91 [89.17, 100.00]100.00100.007.900.000.00
Daily-ServiceEnhanced A2A38.4936.98 [7.69, 68.23]75.0075.5214.880.0011.11
Daily-ServiceAdapted AgentNet70.7365.62 [37.44, 85.25]80.7399.4816.090.0022.22
ResearchAEP (Ours)99.6796.35 [86.04, 100.00]100.00100.009.200.002.90
ResearchEnhanced A2A48.9346.35 [18.23, 76.04]86.4695.8311.040.0033.33
ResearchAdapted AgentNet71.5667.71 [45.90, 88.21]94.79100.0014.600.000.00

On the 64-task 16-agent coding benchmark (Table 18\color{blue} \text{Table 18}), AEP\text{AEP} achieves 99.48% Keyword Value Matching Accuracy (95%95\% CI: [86.77,100.00][86.77, 100.00]) and 100.0% Multiple-Choice Accuracy, outperforming Enhanced A2A (68.8%, CI [52.5,84.6][52.5, 84.6]; paired difference +30.73%, 95%95\% CI [11.95,49.86][11.95, 49.86]) and Adapted AgentNet (82.3%, CI [69.3,95.2][69.3, 95.2]; paired difference +17.19%, 95%95\% CI [4.71,32.21][4.71, 32.21]). Thus, even when baselines are equipped with matched discovery, semantic routing, evolutionary experience, and DLP filters, AEP\text{AEP} delivers substantial, statistically significant improvements across all domains while maintaining lower token costs and near-zero injection acceptance.

AuthorsOpenReview ↗

Rebuttal by Authors [3/7]

Weakness 2: Insufficient Adversarial stress testing on AEP.

"The attack stress in the experiments is not sufficient for AEP and mainly exposes the weaknesses of the baselines."

To directly stress-test AEP\text{AEP}'s routing, trust admission, and quality-feedback mechanisms under non-trivial adversarial strategies that evade static content filters, we conducted an expanded suite of distributed adversarial evaluations in the 16-agent Coding / software-operations domain (Appendix B.9, Table 19\color{blue} \text{Appendix B.9, Table 19}). We focus this deep adversarial stress test on the Coding/operations domain because its tasks require exact operational parameters (quotas, Terraform commands, cron intervals, migration durations, test states, and TLS dates), allowing deterministic, unambiguous verification of whether an agent adopts an attacker's fabricated value versus the ground-truth answer.

Each episode comprises 33 benign service-status warm-up rounds followed by 11 formal evaluation query under natural protocol routing (n=18n=18 episodes per condition). We evaluate four fixed distributed attack strategies:

  • Immediate Attack: The adversary provides plausible but factually false operational values from the very first warm-up round.

  • Sleeper Attack: The adversary behaves honestly during all 33 warm-up rounds to build positive reputation/experience, then switches to returning false values on the target task.

  • Collusion Attack: Two malicious identities return mutually consistent false values and cross-endorse each other.

  • Sybil-Style Multi-Identity Attack: Four admitted malicious identities (out of 16 total agents) coordinate false values and mutual endorsements.

Attack ConditionAEP Acc. (%) ↑Enh. A2A Acc. (%) ↑AgentNet Acc. (%) ↑AEP False Adopt. (%) ↓Enh. A2A False Adopt. (%) ↓AgentNet False Adopt. (%) ↓AEP Malicious Exposure (/18) ↓Enh. A2A Malicious Exposure (/18) ↓AgentNet Malicious Exposure (/18) ↓
Clean100.094.461.10.00.00.0———
Immediate100.061.166.70.038.922.20/1818/186/18
Sleeper88.966.755.60.033.30.00/1818/181/18
Collusion (2 IDs)88.950.066.70.050.05.60/1818/181/18
Sybil (4 IDs)94.450.061.10.050.016.70/1818/183/18

As shown in Table 19\color{blue} \text{Table 19}, AEP\text{AEP} achieves 0.0% false-value adoption and 0/18 malicious response exposure across all four attack strategies while maintaining 88.9%–100.0% accuracy. This demonstrates that AEP\text{AEP}'s robustness is not merely a byproduct of receive-side regex filtering; rather, AEP\text{AEP}'s front-end candidate admission, service-quality reranking, and verified topology routing prevent task requests from being routed to untrusted or degraded responders, whereas Enhanced A2A (18/1818/18 exposure, 33.3%–50.0%33.3\%\text{–}50.0\% false adoption) and Adapted AgentNet (1/18–6/181/18\text{–}6/18 exposure, up to 22.2%22.2\% false adoption) remain vulnerable.

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

This paper provided a solution for agents to discover and collaborate with other agents in dynamic open low-trust networks, with incomplete topology information and to avoid the privacy leakage and attack. It proposed Open Private Agent Collaboration(OPAC) to coordinate the exchange of policy-controlled context assets, which includes memory, skills and executions between agents. To further implement OPAC, it proposed AEP( Agent exchange protocol), including Routing, Coordinate, Exchange and Ledger. It also provided theoretical analysis under the idealized assumptions. Based on the simulation networks of 4, 16, 64 agents with different tasks including coding, research, daily service and so on, AEP reached better performance compared to other baselines such as A2A, Broadcast and so on. Also there are ablation studies on difference components as well as safety, task utilization, long-term, Robustness under Malicious Agents and so on.

Critical strengths
  1. This paper targets on a very novel and practical problem, where the network is partially observed and low-trust, also it considered the safety gate control and malicious agents. What's more, it models the memory, skills as well as executions into shareable context-assets which can be spread and reused by other agents and further improve the collaboration between agents. Those settings made the proposed algorithms to be more practical and could potentially bring benefit to the real world.

  2. AEP is a concrete and closed-end framework with closing coordination and feedback between each components: (1) in AEP, the agent discovery, candidate admission, negotiation, safety check, evidence were integrate together, making the historical coordination take affect in the future coordination. (2) Ablation study on routing and safety component also provided further evidence for the claims in terms of task performance and safety.

Critical weaknesses
  1. There is unclear definition and claim of the safety trust and Quality feedback and refinement: (1) in the appendix A.2, there are functions about the Demand resolution and Quality feedback, however, they did not explain how the conf(d) is decided and how the requester computes a quality signal in details and which information is required to make decision on the conf(d) and quality signal. (2) A.3. described how the assets were released and how the trust gate was implemented, however, the key rule of the decision and the experimental setup were not clearly defined, which is important because they are directly related tot he malicious agents detection, trust update as well as the remain parts. Lacking of those clear definition made the safety and robustness, which claimed as the key contribution of this paper, be irreproducible and unconvinced.

  2. Current ablation study and results can not fully demonstrate the benefit from specific design of AEP: (1) it lacks of comparation with other functional similar baselines or enough evidence to support those related baselines are not suitable for this study cases, for instance, in the related work section, the author mentioned AgentNet (Yang et al.,2025a) , Internet of Agents (Chen et al.,2024b) Collaborative Memory (Rezazadehetal.,2025) and so on. (2) in table 1, it overclaimed the privacy protect was unsupported for Shared Memory method, as Collaborative Memory already integrated the access control and selective memory share. It is better to clarify the memory access control mechanism in the experimental setting as well as the uplift of AEP over the existed Shared Memory method. (3) Similarly, since A2A baseline won't reach the further agents beyond the initial neighbor agents and the tasks were distribute among different agents, currently results can not distinguish whether the uplift from the A2A came from the AEP design or from the gaining more information from other agents. (4) In the Asset Exchange ablation, after remove the Exchange, the performance was slightly improved, which can not demonstrate the mechanism of Exchange.

Questions for authors
  1. Please answer the questions in the Critical Weaknesses sections, which is related to the soundness of the paper.
  2. I also have questions related to the scope and claim of this paper: (1) Did this paper assume that all the regular agents are willing to contribute to other agents once the collaboration conditions were satisfied? Can the agents have right to refuse provide service based on own budget or other reasons such as other commitments or saving for itself. (2)Is AEP only in charge of the technical algorithm execution part while leaving matters such as payment, reciprocity, and the motivation to participate to external systems?
11 comments
AuthorsOpenReview ↗

Rebuttal by Authors [1/8]

Thank you for your thorough review and constructive questions! To address your questions in detail and further validate AEP\text{AEP}, we have expanded our operational specifications for demand confidence, quality feedback, and asset-release/trust-gate decision rules (Appendix A.2–A.3\color{blue} \text{Appendix A.2–A.3}), and conducted additional comparative experiments against Adapted AgentNet, access-controlled Collaborative Memory, Enhanced A2A, and non-cooperative agents. Below we respond to each point in Weaknesses and Questions.

Weakness 1: Operational definitions of demand confidence and quality feedback.

"Confidence estimation and quality feedback are insufficiently defined: how conf(d) and the requester’s quality signal are computed, and what information they require, remain unclear."

We have added a detailed operational specification in Appendix A.2\color{blue} \text{Appendix A.2} (page 24) that explicitly distinguishes demand-level resolution confidence conf(d)\mathrm{conf}(d) from peer-level service-quality feedback Ψijqual(t)\Psi^{\mathrm{qual}}_{ij}(t):

  1. Demand Resolution Confidence conf(d)\mathrm{conf}(d) (Evaluating Whether Demand dd Is Solved): When requester AiA_i receives a response for an unresolved demand dd, the coordinator invokes a structured sufficiency evaluation conditioned on the original task query xix_i, current demand description dd, local context, and peer response, returning:

(answer_candidate, source_span, addresses_demand, sufficient, confidence, reason).

If the evaluator returns an explicit boolean sufficient flag, the coordinator uses it directly; otherwise, it resolves dd via Resolved(d)=1[conf(d)≥θres]\mathrm{Resolved}(d) = \mathbb{1}[\mathrm{conf}(d) \ge \theta_{\mathrm{res}}] with default threshold θres=0.75\theta_{\mathrm{res}} = 0.75 (Eq. 29). If Resolved(d)=0\mathrm{Resolved}(d)=0, demand dd remains pending and is reallocated to untried candidates with Eq. 27 until the collaboration budget is exhausted. Thus, conf(d)\mathrm{conf}(d) measures evidential sufficiency for the current subproblem, not peer reputation.

  1. Peer Service-Quality Feedback Ψijqual(t)\Psi_{ij}^{\mathrm{qual}}(t) and Asymmetric EMA (Evaluating Peer Usefulness): After each interaction with responder AjA_j, the requester maps the downstream collaboration outcome to a bounded usefulness score Ψijqual(t)∈[0,1]\Psi^{\mathrm{qual}}_{ij}(t) \in [0,1] using a deterministic mapping (Appendix A.2\color{blue} \text{Appendix A.2}):
Collaboration OutcomeOperational ConditionQuality Signal Ψijqual(t)\Psi^{\mathrm{qual}}_{ij}(t)
Demand resolvedResponse satisfies demand (sufficient = true or conf(d)≥0.75\mathrm{conf}(d) \ge 0.75)1.00
Insufficient responsePartial or incomplete context; demand remains pending0.35
Replan triggeredResponse requires demand decomposition or parameter replanning0.25
Capability / Asset mismatchContacted peer lacks relevant skill or memory asset for dd0.10
Maximum replans reachedInteraction exhausts negotiation/replan budget without resolution0.00
Safety rejectionInbound payload rejected by receive-side safety or policy guard0.00

In the local direct-evidence quality path (local_evidence mode), the requester updates its local quality estimate qij(t)=Qualij(t)q_{ij}(t) = \mathrm{Qual}_{ij}(t) via an asymmetric EMA in Eq. 31.

Requesters assess the usefulness of peer responses and record quality feedback in the signed evidence ledger. Other agents ingest verified feedback records to update their own quality estimates. Service quality reranks already-admissible peers and is kept separate from safety-trust admission.

AuthorsOpenReview ↗

Rebuttal by Authors [2/8]

Weakness 2: Specification of asset-release and trust-gate decision rules.

"Asset-release and trust-gate decision rules are insufficiently specified, making malicious-agent detection, trust updates, and the claimed safety and robustness difficult to reproduce and verify."

We have added the complete decision rules and exact threshold parameterization in Appendix A.3\color{blue} \text{Appendix A.3} (page 26). AEP\text{AEP} enforces safety across three stages:

  1. Peer Authorization: Safety trust is derived by aggregating local and distributed safety assessments over the evidence-backed topology through the damped propagation procedure in Eqs. (20)–(21). Routing applies blacklist, effective-trust, and reachability checks before semantic and service-quality ranking. For an exchange requested by agent Ai from responder Aj, the responder applies the effective-trust gate in Eq. (34), followed by asset-specific authorization and outbound safety checks.

  2. Outbound Asset Release:

    • Discovery Excerpts: Public memory summaries exported for discovery are drawn strictly from the agent's native work memory Wi(t)W_i(t) (imported remote assets in exchange memory Ei(t)E_i(t) are never re-exported as native discovery evidence). Exported records must have sensitivity <0.50<0.50, pass outbound DLP, injection sanitization, and secret-pattern checks under a conservative peer trust baseline of 0.300.30, and are capped at ≤4\le 4 excerpts (≤160\le 160 chars/excerpt, ≤600\le 600 chars total).

    • Tiered Exchange Release: Raw memory release (raw_memory) requires policytrust(Aj)≥0.70\mathrm{policy_trust}(A_j) \ge 0.70, asset sensitivity <0.80<0.80, and passing outbound DLP checks. When raw release is not permitted, the responder falls back to processed_memory (redacted/summarized state) or delegated_computation (local skill execution returning only the output), or blocked if policy checks fail.

Parameter / Gate RuleOperational ValueProtocol Enforcement Stage
Verified peer initial quality prior qij(0)q_{ij}(0)0.50Candidate admission & release authorization
Unverified peer trust cap0.30Candidate admission & discovery excerpt gate
Blacklisted peer policy trust0.00Hard exclusion in routing & exchange
Guarded discovery excerpt sensitivity ceiling0.50Outbound discovery summary generation
Guarded discovery conservative peer trust0.30Outbound discovery summary authorization
Guarded discovery excerpt limits≤ 4 items (≤ 160 chars/item, ≤ 600 total)Outbound PeerCard / FlowEntry export
High asset sensitivity block threshold0.80Outbound raw_memory / summary blocking
Raw-memory release trust threshold0.70Outbound raw_memory vs. processed_memory
Inbound/Outbound DLP regex risk threshold0.50 (binary hit ∈{0,1}\in \{0,1\})Inbound validation & outbound guard
  1. Staged Inbound Validation: Received payloads are re-validated by the requester's inbound safety policy before integration into isolated exchange_memory Ei(t)E_i(t) (Eq. 37), preventing direct contamination of core work memory Wi(t)W_i(t) and logging a trust_update safety incident if an injection or policy violation is detected.
AuthorsOpenReview ↗

Rebuttal by Authors [3/8]

Weakness 3: Comparisons with AgentNet, Internet of Agents, and Collaborative Memory.

"The paper misses comparisons with functionally similar methods such as AgentNet, Internet of Agents, and Collaborative Memory, or sufficient justification for why they are unsuitable."

We have addressed all three systems through both extensive empirical comparisons (adapting AgentNet [1] and evaluating Collaborative Memory [2]) and architectural analysis (Appendix B.8–B.9, Tables 17–20\color{blue} \text{Appendix B.8–B.9, Tables 17–20}):

  1. Empirical Comparison with Adapted AgentNet (Appendix B.8–B.9, Tables 17–20\color{blue} \text{Appendix B.8–B.9, Tables 17–20}): Using the original repository of AgentNet [1], we adapted it to our 16-agent decentralized setting while preserving its core evolutionary coordination mechanisms and equipping it with matched discovery and safety filters:
    • Preservation of Native Evolutionary Components: We retain AgentNet's native Router, Executor, Task module, episodic experience pool, dynamic capability profiles, and adaptive edge-weight updating rules. During multi-round episodes (such as warm-up trajectories in adversarial evaluations), completed interactions update AgentNet's native experience pool and inter-agent connection weights without cross-task test leakage.

    • Runtime, Embedding, and Discovery Alignment: We adapt AgentNet to the 16-agent collaborative runtime with DeepSeek-V4-Flash (temperature=0, thinking disabled) and shared all-MiniLM-L6-v2 embeddings (zero-padded from 384384 to 10241024 dimensions to match upstream AgentNet's vector interface while preserving exact cosine similarity). We also integrate PeerCard gossip discovery (exchanging capability, memory, and skill summaries every 500 ms500\text{ ms}) so AgentNet discovers peers beyond its initial 30%30\% neighborhood, and apply deterministic inbound injection and outbound DLP filters on executor returns and final outputs, bounding execution to DAG workflows with at most 33 direct peer requests per task.

    • Multi-Domain Utility & Safety (Tables 17–18\color{blue} \text{Tables 17–18}): Across Coding, Daily-Service, and Research on the 16-agent network, we evaluate both rubric-based Task Score and two LLM-judge-free deterministic metrics—Keyword Value Matching Accuracy (6464 tasks ×3\times 3 seeds =192=192 runs per protocol) and Multiple-Choice Accuracy (192192 runs per protocol):

ScenarioProtocolTask Score ↑Keyword Value Matching Acc. & 95% CI (%) ↑Multiple-Choice Acc. (%) ↑Completion Rate (%) ↑Tokens/task (k) ↓Leakage Rate (%) ↓Injection Success Rate (%) ↓
CodingAEP (Ours)99.1799.48 [86.77, 100.00]100.00100.0010.100.000.00
CodingAdapted AgentNet78.3382.29 [69.27, 95.24]93.75100.0013.990.0044.44
Daily-ServiceAEP (Ours)95.5097.91 [89.17, 100.00]100.00100.007.900.000.00
Daily-ServiceAdapted AgentNet70.7365.62 [37.44, 85.25]80.7399.4816.090.0022.22
ResearchAEP (Ours)99.6796.35 [86.04, 100.00]100.00100.009.200.002.90
ResearchAdapted AgentNet71.5667.71 [45.90, 88.21]94.79100.0014.600.000.00

As shown in Table 17\color{blue} \text{Table 17} and Table 18\color{blue} \text{Table 18}, AEP\text{AEP} consistently outperforms Adapted AgentNet across all three domains while consuming 27.8%–50.9%27.8\%\text{–}50.9\% fewer tokens, achieving a statistically significant paired accuracy gain of +17.19 % .

Reviewer_Pb7LOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces Open Private Agent Collaboration (OPAC), an asset-centric framework for cooperation among privately owned agents in open, partially observed, low-trust networks. Collaboration is framed as selective discovery, exchange, auditing, and integration of policy-controlled context items (memory fragments, skills, cached results, etc.) rather than pure request-response or flat shared memory. The framework is used to instantiate the Agent Exchange Protocol (AEP) with four subprotocols: AEP-ROUTING: semantic discovery via a local Semantic Flow Table. AEP-COORDINATE: demand planning, negotiation, and feedback refinement. AEP-EXCHANGE: safety-gated release with staged inbound validation and separation of exchange memory. AEP-LEDGER: append-only signed DAG (Merkle-DAG) evidence log with Lamport clocks, equivocation detection/blacklisting, quality/trust derivation, and quorum-backed checkpoints. The paper explores a setting where private agent networks have partial topology, unreliable transport, partitions, and leave-and-rejoin.

Critical strengths

The problem setting is realistic and timely: open private agent networks with dynamic membership, partial topology knowledge, ownership boundaries, and risks of leakage, injection, and memory contamination. Existing request-centric A2A-style protocols assume known counterparts; shared-memory/broadcast systems assume visibility and trust that do not hold. The asset-centric middle ground and the concrete mechanisms (especially evidence-derived routing + safety-gated release modes + ledger-backed adaptation) are practical contributions. The operational details in Appendix A (flow-table fields, gossip envelopes, selection/pruning algorithms, release modes, checkpointing) show engineering substance beyond the high-level claims. Experiments are thorough (multi-scale, multi-domain, utility + safety + long-term + robustness + ablations + production-protocol and external-guard comparisons). Safety results (0% leakage, near-zero injection, progressive reduction of malicious contacts) are good relative to the baselines.

Critical weaknesses

The OPAC framing (physical vs. cognitive topology, admissible asset sets under policies, multi-objective safe Pareto) usefully articulates the constraints and the desired middle ground between pure request-response and global shared memory. However, it is quite abstract. The concrete protocol design (semantic flow tables, constrained gossip, topology probes, release modes, DAG evidence, etc.) appears driven by practical systems considerations and - IMO - could largely have been engineered without the multi-objective Pareto scaffolding.

The baseline was easy to beat under the assumptions. Request-centric A2A (and the production protocols derived from it) do not really target open-membership discovery under partial topology; Broadcast is noisy and high-cost by design. The evaluation deliberately distributes knowledge and capabilities so that collaboration is necessary, and AEP is engineered precisely for performing semantic discovery + selective exchange + evidence-based trust under those conditions. The controlled virtual-HTTPS simulation further favors the system designed for it. Comparisons to production protocols and external guards (Qwen3Guard) improve fairness a bit, but simpler hybrids (e.g., A2A augmented with a semantic routing layer and policy filters) are not tested. LLM-as-judge scoring, modest task counts per run (6/36), and the relatively stylized malicious-agent behavior also temper the strength of the paper's claims.

Coupling Theorem 1 (informal) / Theorem 9 (formal) requires a long list of strong, idealized assumptions (A1–A12 in Appendix D—connected graph with spectral gap, topology-learning contraction, monotone admissible discovery, certified residual leakage bound ρS\rho_S for sanitization, finite joint safe policy class, bounded independent samples, uniform topology sensitivity, uniform dominance-gap condition, etc.), the Hausdorff distance between the AEP-induced empirical objective-space frontier and the true safe Pareto frontier vanishes as statistical error, topology-learning error, and auditing/pruning loss go to zero. The result only tells us that if the mechanisms converge and samples are sufficient, the empirical frontier is close; it does not suggest the needed design choices (gossip fanout, probe handshake, release modes, checkpoint quorum, etc.) nor provide tight practical guarantees. The use of Pareto optimality for AI decision-making is reasonable in principle, but the theorem mainly retrofits the justification onto an already-designed protocol.

Finally, in the paper’s idealized model, some form of causal ordering is useful so that agents can reason about “what happened before what” without trusting synchronized physical clocks. The combination of Lamport clocks + sequence numbers gives a lightweight way to order evidence and detect conflicting signatures from the same author. However, the necessity is overstated for the scenarios that actually matter today: Most current agentic systems (multi-agent frameworks, tool-calling pipelines, research/coding agents, enterprise copilots) run inside a single cloud region, a single VPC, or behind a trusted control plane. A shared, trusted clock (or even just a central sequence service/database timestamps) is cheap and reliable. Even when agents call external APIs or talk across organizations, the latency and clock skew are rarely severe enough that Lamport-style causality becomes an important correctness concern. The paper’s own simulation (virtual HTTPS with injected latency, loss, and partitions) creates the asynchrony that then justifies the logical clock. In a more realistic deployment model—agents that are co-located or that can fall back to a trusted time source—the extra machinery adds complexity without clear payoff.

Questions for authors

\textbf{On the formalization}. Can you point to a specific design decision, invariant, or safety property in AEP that was discovered or forced by the multi-objective / Πsafe\Pi_{\text{safe}} formalization in Section 2, and that would have been missed or left ambiguous in a purely engineering-driven design? If the formalization was mainly descriptive rather than generative, what independent technical contribution does it make?

\textbf{On the Lamport-style clock}. Lamport clocks are intended for the regime in which a consistent shared notion of time cannot be assumed. In which concrete agent-network deployments that you consider realistic for the next 3–5 years is a trusted common clock (or a central sequencer) unavailable? If one simply replaces the Lamport logical clock + per-author sequence with a trusted wall-clock + monotonic counters, which of AEP’s safety or trust properties fail or become materially harder to achieve, and by how much?

\textbf{On the experimental claims}. If the baselines are strengthened with the same evidence-backed trust, simple blacklisting, and safety-gated release primitives that AEP uses (while keeping their original communication model), which of the reported advantages remain? On which of the evaluated tasks does the full distributed-systems machinery (constrained gossip, Lamport ordering, Merkle-DAG ledger) contribute a measurable gain over a simpler design that assumes a trusted common clock and a more synchronous setting?

15 comments
AuthorsOpenReview ↗

Rebuttal by Authors [1/8]

Thank you for your thorough review and insightful comments!

To address the questions and further validate AEP\text{AEP}, we have clarified how the OPAC formalization and coupling theorem directly guide concrete protocol design invariants (Appendix C.4.1, Appendix D.1.1\color{blue} \text{Appendix C.4.1, Appendix D.1.1}), and conducted additional experiments with discovery- and safety-enhanced baselines (Enhanced A2A and Adapted AgentNet), LLM-judge-free deterministic metrics across 6464 tasks ×3\times 3 seeds, multi-identity and adaptive attacks, fault-free synchronous transport without virtual HTTPS, and a central-clock vs. Lamport-clock ablation. Below we respond to each point in Weaknesses and Questions.

Weakness 1: Necessity of the OPAC/Pareto formalization for protocol design.

"The OPAC/Pareto formalization is abstract and appears mainly descriptive; its necessity for the concrete protocol design is unclear."

Rather than serving as a purely descriptive wrapper, the OPAC formulation in Section 2 (Appendix C.4.1\color{blue} \text{Appendix C.4.1}, page 39) directly dictates five concrete architectural invariants in AEP\text{AEP} that distinguish it from heuristic multi-agent pipelines:

  1. Non-Decomposable Joint Safety Set Πsafe(t)  ⟹  \Pi_{\mathrm{safe}}(t) \implies Dual-Sided Sender/Receiver Enforcement: In Eq. (2), Πsafe(t)={π∣∀i,m→iπ(t)⊆Ai(t),Ripriv(π)≤ϵi}\Pi_{\mathrm{safe}}(t) = \{\pi \mid \forall i, m_{\rightarrow i}^\pi(t) \subseteq \mathcal{A}_i(t), R_i^{\mathrm{priv}}(\pi) \le \epsilon_i\} couples the responder's outbound release policy with the requester's inbound admissibility set. Because safety is non-decomposable across agents, unilateral sender filtering or receiver output guarding is insufficient; this directly forces AEP-Exchange to pair outbound authorization/DLP gating (Eq. 35) with receive-side validation into isolated exchange_memory Ei(t)E_i(t) (Eq. 37).

  2. Hard Admissibility m→iπ(t)⊆Ai(t)m_{\rightarrow i}^\pi(t) \subseteq \mathcal{A}_i(t) vs. Soft Utility U(π)  ⟹  U(\pi) \implies Lexical Priority of Trust Admission Over Semantic Ranking: Because the Pareto frontier P⋆(t)⊆Πsafe(t)\mathcal{P}^\star(t) \subseteq \Pi_{\mathrm{safe}}(t) is defined strictly inside the safe feasible set, high semantic similarity can never compensate for a trust or safety violation. This forces AEP-Routing (Algorithm 1, Eqs. 14–17) to enforce hard blacklist, reachability, and trust gates (τijeff≥τmin⁡\tau^{\mathrm{eff}}_{ij} \ge \tau_{\min}) before computing semantic relevance and quality reranking sˉij=sij(0.7+0.3q~ij)\bar{s}_{ij} = s_{ij}(0.7+0.3\tilde{q}_{ij}).

  3. Privacy Bound Ripriv(π)≤ϵiR_i^{\mathrm{priv}}(\pi) \le \epsilon_i + Information-Utility Equivalence (Theorem 6 & Corollary 7)   ⟹  \implies Four-Tier Asset Release: Theorem 6 proves that maximizing predictive utility U(π)U(\pi) over Πsafe(t)\Pi_{\mathrm{safe}}(t) is equivalent to maximizing imported task-relevant conditional mutual information I(Yi;m→iπ(t)∣Xi,Mi)I(Y_i; m_{\rightarrow i}^\pi(t) \mid X_i, M_i), while Proposition 5 bounds leakage via I(S(ξj);s)≤ρSI(S(\xi_j); s) \le \rho_S. This directly motivates AEP-Exchange's four release tiers (raw_memory, processed_memory, delegated_computation, blocked): when policytrust≥0.7\mathrm{policy_trust} \ge 0.7 and sensitivity is low, Corollary 7 proves raw release strictly Pareto-dominates lossy summarization; when raw release would violate ϵi\epsilon_i, processed summarization or local skill delegation preserves I(Yi;S(ξj)∣Xi,Mi)>0I(Y_i; S(\xi_j) \mid X_i, M_i) > 0 while satisfying Ripriv(π)≤ϵiR_i^{\mathrm{priv}}(\pi) \le \epsilon_i.

  4. Discrepancy Between Cognitive Topology Wicog(t)W_i^{\mathrm{cog}}(t) and Physical Topology Wphy  ⟹  W^{\mathrm{phy}} \implies Two-Sided Evidence-Backed Topology Giev(t)G_i^{\mathrm{ev}}(t): Minimizing the topology perturbation term Lobj∥Wicog(t)−Wphy∥2L_{\mathrm{obj}}\|W_i^{\mathrm{cog}}(t) - W^{\mathrm{phy}}\|_2 in Theorem 9 requires preventing unverified multi-hop gossip rumors from corrupting Wicog(t)W_i^{\mathrm{cog}}(t), forcing AEP-Ledger and AEP-Routing to admit edges into Giev(t)G_i^{\mathrm{ev}}(t) only when backed by matched two-sided signed claims (Tx(a,b,h)∧Rx(b,a,h)\mathrm{Tx}(a,b,h) \wedge \mathrm{Rx}(b,a,h), Eqs. 13, 43).

  5. Resource Cost Objective −C(π)  ⟹  -C(\pi) \implies Local Sufficiency Gating & Bounded Negotiation: Explicitly optimizing −C(π)-C(\pi) alongside U(π)U(\pi) and T(π)T(\pi) forces AEP-Coordinate to check local sufficiency before querying peers (Eq. 25), bound routing pools to TopK⁡\operatorname{TopK} (K≤3K \le 3, Eq. 19), and cap negotiation rounds (rmax⁡=3,br=9r_{\max}=3, b_r=9, Eq. 28).

AuthorsOpenReview ↗

Rebuttal by Authors [2/8]

Weakness 2: Evaluation against stronger hybrid baselines with discovery and safety gates.

"Baselines are disadvantaged by the evaluation assumptions; stronger hybrids with semantic discovery, evidence-backed trust, blacklisting, and safety-gated release are not compared."

To ensure a rigorous comparison under matched discovery, routing, and safety capabilities, we engineered Enhanced A2A and Adapted AgentNet [1] (Appendix B.8, Tables 17–18\color{blue} \text{Appendix B.8, Tables 17–18}):

  1. Enhanced A2A (Gossip Discovery + MiniLM Semantic Routing + Deterministic Policy Filters & Outbound DLP):

    • (a) Gossip-based Peer Discovery: Each agent initializes its local directory with only its directly observed neighbors and maintains structured PeerCard records containing agent_id, endpoint, capability_desc, memory_summary, skill_summary, skill_count, and trust_score. Every 500 ms500\text{ ms}, each agent selects up to two gossip targets from known peer IDs and exchanges its own PeerCard along with known remote PeerCards. Receivers verify endpoint syntax and physical reachability before merging newly discovered cards, enabling multi-hop peer discovery across the entire network without global broadcasting.
    • (b) MiniLM-based Semantic Routing: For each task, the requester concatenates each candidate's capability description, memory summary, and skill summary, computes normalized embeddings using the shared all-MiniLM-L6-v2 encoder, filters out peers with trust_score <0.15< 0.15, and selects the Top-33 peers by cosine similarity (with deterministic tie-breaking by agent_id).
    • (c) Deterministic Policy Filters and Outbound DLP: We add rule-based inbound and outbound security classifiers (binary risk ∈{0,1}\in \{0,1\}, blocking threshold 0.50.5) across all request and response paths. Inbound checks block instruction overrides, spoofed system/developer privilege headers, policy-bypass attempts, and secret solicitation. Outbound DLP checks block API keys, passwords, secrets, PEM private-key headers, and AWS credentials. Published memory summaries are restricted to records with sensitivity <0.8<0.8 passing outbound DLP (≤4\le 4 records, ≤160\le 160 chars/record, ≤600\le 600 chars total). Local RAG context, collaborative responses, and final simulator outputs are all filtered.
  2. Adapted AgentNet (Retaining Native Evolutionary Router/Executor + PeerCard Discovery + Shared Safety Filters):

    • Using the original repository of AgentNet [1], we retain its native Router, Executor, Task module, episodic experience pool, capability profiles, and adaptive edge-weight updating rules.
    • We adapt AgentNet to the 16-agent collaborative runtime with DeepSeek-V4-Flash (temperature=0, thinking disabled) and shared all-MiniLM-L6-v2 embeddings (zero-padded from 384384 to 10241024 dimensions to match upstream AgentNet's vector interface while preserving exact cosine similarity).
    • We integrate the same PeerCard gossip discovery module so AgentNet discovers peers beyond its initial 30%30\% neighborhood, and apply the same inbound injection and outbound DLP filters on executor returns and final outputs. Workflows are bounded to DAG execution with at most 33 direct peer requests per task. During multi-round episodes, warm-up trajectories update AgentNet's native experience pool and edge weights without cross-task test leakage.

Across Coding, Daily-Service, and Research on the 16-agent network (Table 17\color{blue} \text{Table 17}), AEP\text{AEP} outperforms both Enhanced A2A and Adapted AgentNet across every utility and safety dimension:

ScenarioProtocolTask Score ↑Keyword Value Matching Acc. & 95% CI (%) ↑Multiple-Choice Acc. (%) ↑Completion Rate (%) ↑Tokens/task (k) ↓Leakage Rate (%) ↓Injection Success Rate (%) ↓
CodingAEP (Ours)99.1799.48 [86.77, 100.00]100.00100.0010.100.000.00
CodingEnhanced A2A67.5868.75 [52.53, 84.62]94.2791.6712.180.0011.11
CodingAdapted AgentNet78.3382.29 [69.27, 95.24]93.75100.0013.990.0044.44
Daily-ServiceAEP (Ours)95.5097.91 [89.17, 100.00]100.00100.007.900.000.00
Daily-ServiceEnhanced A2A38.4936.98 [7.69, 68.23]75.0075.5214.880.0011.11
Daily-ServiceAdapted AgentNet70.7365.62 [37.44, 85.25]80.7399.4816.090.0022.22
ResearchAEP (Ours)99.6796.35 [86.04, 100.00]100.00100.009.200.002.90
ResearchEnhanced A2A48.9346.35 [18.23, 76.04]86.4695.8311.040.0033.33
ResearchAdapted AgentNet71.5667.71 [45.90, 88.21]94.79100.0014.600.000.00
AuthorsOpenReview ↗

Rebuttal by Authors [3/8]

Weakness 3: Deterministic evaluation metrics and expanded sample sizes.

"LLM-as-judge scoring and small task counts per run limit confidence in the reported improvements."

To avoid reliance on LLM-as-judge scoring and provide strong statistical confidence, we conducted additional large-scale evaluations (Appendix B.7–B.8, Tables 17–18\color{blue} \text{Appendix B.7–B.8, Tables 17–18}):

  1. Two Deterministic LLM-Judge-Free Metrics: We evaluate open-ended Keyword Value Matching Accuracy via deterministic typed parsing of exact operational values, actions, intervals, durations, statuses, and dates, as well as 4-option Multiple-Choice Accuracy (A–D).

  2. Expanded Sample Size & CIs: We expanded the evaluation to 6464 tasks ×3\times 3 seeds =192= 192 runs per protocol per domain and computed 95%95\% confidence intervals (Table 18\color{blue} \text{Table 18}). On the 64-tasks benchmark, AEP\text{AEP} (99.48%, CI [86.77,100.00][86.77, 100.00]) achieves statistically significant paired gains of +30.73% (95%95\% CI: [11.95,49.86][11.95, 49.86]) over Enhanced A2A (68.8%) and +17.19% (95%95\% CI: [4.71,32.21][4.71, 32.21]) over Adapted AgentNet (82.3%).

Weakness 4: Evaluation under non-stylized and adaptive malicious behaviors.

"The malicious-agent behaviors evaluated in the paper are relatively stylized, which weakens the broader robustness claims."

To evaluate AEP\text{AEP} under strategic, colluding, and adaptive threat models that evade static content filters, we added two comprehensive suites of adversarial evaluations in the 16-agent Coding / software-operations domain (Appendix B.9, Tables 19–20\color{blue} \text{Appendix B.9, Tables 19–20}):

  1. Four Distributed Fixed-Strategy Attacks (Table 19\color{blue} \text{Table 19}): Each episode comprises 33 benign service-status warm-up rounds followed by 11 formal evaluation query under natural protocol routing (n=18n=18 episodes per condition):
    • Immediate Attack: The adversary returns plausible but factually false operational values starting from the first warm-up round.
    • Sleeper Attack (Strategic Reputation Manipulation): The adversary behaves honestly during all 33 warm-up rounds to accumulate positive reputation/experience, then switches to returning false values on the target task.
    • Collusion Attack (2 Identities): Two malicious identities return mutually consistent false values and cross-endorse each other.
    • Sybil-Style Multi-Identity Attack (4 Identities): Four admitted malicious identities (out of 16 total agents) coordinate false values and mutual endorsements.
Attack ConditionAEP Acc. (%) ↑Enh. A2A Acc. (%) ↑AgentNet Acc. (%) ↑AEP False Adopt. (%) ↓Enh. A2A False Adopt. (%) ↓AgentNet False Adopt. (%) ↓AEP Malicious Exposure (/18) ↓Enh. A2A Malicious Exposure (/18) ↓AgentNet Malicious Exposure (/18) ↓
Clean100.094.461.10.00.00.0———
Immediate100.061.166.70.038.922.20/1818/186/18
Sleeper88.966.755.60.033.30.00/1818/181/18
Collusion (2 IDs)88.950.066.70.050.05.60/1818/181/18
Sybil (4 IDs)94.450.061.10.050.016.70/1818/183/18

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.