acceptodds
Withdrawn from ICLR 2027

Cross-Principal Communication Between Agent Delegates

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Agents increasingly act as delegates with access to their owners’ private information and tools. Communication across owners, or principals, enables cooperation but must respect constraints on disclosure and state changes, creating a security–utility trade-off. Studying that trade-off requires a framework in which communication policies and executable permissions are controlled together, and in which messages are connected to the resource access and state changes they cause. We formulate cross-principal delegate communication and introduce such a framework: SharedOS governs execution under explicit permissions, while SharedEval lets researchers define environment-specific tasks, authorization labels and outcome criteria reflecting owners’ expectations. We instantiate it with two datasets, bilateral PAIR and network NET. In PAIR, more explicit, relationship-aware policies reduce disclosure, their utility costs depend on requester entitlements, and requester relationships shape both disclosure and over-refusal. Re-asking a refused request under new framings almost never changes the delegate’s decision, although a multi-turn session exposes more of the protected material than a single request does. In NET the same ladder does not hold: the holder’s own relationship-aware document gives back more than the generic caution gained, because a forbidden edge is one fact inside a category that the same relationship permits. Together, our formulation, research infrastructure and empirical findings provide a foundation for studying cross-principal communication between agent delegates.

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

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_kXnWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies cross-principal information sharing among LLM agent delegates. It introduces SharedOS, an execution layer that enforces owner-issued grants, and SharedEval, which labels what each requester is entitled to receive. The framework is used to evaluate responder delegates in two scenarios: a bilateral setting (PAIR) and a multi-party network (NET).

Critical strengths
  • The topic is important. As agents increasingly act on behalf of different owners, how they handle information sharing across principals is a practical and growing concern.

  • In particular, the multi-party setting of NET is relatively well defined. Its design could support future multi-agent evaluation in information-sharing contexts.

Critical weaknesses
  • The motivation is hard to follow. To evaluate multi-agent systems, one could build a single shared sandbox environment and drive each agent through its own harness SDK. Existing sandboxes already support this. Thus, it is initially unclear why SharedOS is needed. The need is clearer if each agent brings its own sandbox and permission system, in which case a common governing layer such as SharedOS is necessary. The paper would benefit from clarifying this.

  • There are several internal inconsistencies that the authors should clarify (see Questions).

  • Many design decisions are not justified, which makes the work hard to understand and evaluate. For example, both PAIR and NET are constructed by the authors themselves, but the paper does not explain the basis for their design or why the tasks were designed this way. It is also unclear why requester reframing is treated as an experimental variable in PAIR. Moreover, the requester-side variation in PAIR is limited, and it does not necessarily support the general conclusion that "the boundary is set on the responder's side."

Questions for authors
  1. PAIR and NET seems to be designed by the authors themselves. How were the settings of these two scenarios designed, and how was their quality ensured? For example, how can the authors ensure that these scenarios are representative for testing cross-principal delegation, specifically information sharing?

  2. In PAIR, some gold labels in relationship-v2 conflict with the agent-facing D2 policy in appendix d.9.3. For example, 52 of 90 sampled D2 leaks involve items that the category contract treats as public, and 33 of the executive delegate's authorized work actions are withheld by the labels but granted by the D2 text. Why do they conflict? Since the policy and the judgment criteria disagree, do these conflicts undermine the validity of the results? How much do they affect the reported numbers and the findings?

  3. How was NET executed? Sec 4.3 reports runs with 60 agents, while appendix c mentions that NET is not a production execution target. and that the 60-principal v2 world has "no integrated runner." While the abstract presents NET findings as a main result.

  4. The NET ladder result rests on 15 of 136 forbidden edges (the intersection that every policy rung delivered), and the D1–D2 difference in Lw corresponds to a single edge. Why did 4 of the 19 primary edges fail to deliver under some rung, and is this attrition independent of the policy? Can the authors report results on all 19 (or all 49 scorable) edges, treating failures explicitly?

Reviewer_LjtHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper produces and evaluates SharedOS to governs agentic execution under explicit permissions, and SharedEval to constraint evaluation environments.

Critical strengths
  • SharedOS/SharedEval is a plausible research infrastructure. The distinction between governed execution and scenario-specific evaluation is sensible, and the PAIR/NET instances demonstrate different experimental axes.

  • The length-matched generic control, framing controls, cross-model requester swaps, and Codex CLI harness check are useful. The paper does not only compare a long policy to no policy.

Critical weaknesses

The main text is extremely dense and uses many unexplained shorthand. The NET reversal is conceptually the most interesting result, but it is evaluated on little data. The current instrument agreement on NET is also low. The NET world is carefully constructed but small in matched edges.

Questions for authors

The paper says source includes frozen inputs and manifests, but it is not clear whether the full dataset, policies, and harness will be publicly released. For a framework paper, artifact availability is essential.

Can you expand NET to more matched edges and holders?

"Limitations and next steps" section seems rather awkward dump of AI text, hard to understand

Reviewer_q9arOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper contributes an execution and evaluation framework for settings with multiple agents that act as delegates for humans (inheriting access privileges to private data and resources) and need to co-operate while respecting constraints on disclosure of private/sensitive information. It also contributes two related benchmark datasets. This is an important and practical problem and the intent seems to be that researchers should be able to vary components like ommunication policies, execution rights and agent behavior separately for experimentation and evaluation, which is definitely an important criterion.

However the paper suffers from very poor writing and exposition which makes it difficult to understand how the claims are being implemented in the framework. Similarly the motivation and thinking between design of PAIR and NET and comparison to existing benchmarks is lacking.

Critical strengths

None

Critical weaknesses

Poor writing and comparison to existing work.

Questions for authors

None

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

Related papers

Discussion (0)

Sign in to comment.