acceptodds
Withdrawn from ICLR 2027

PolicyFork: Task Success Can Hide Policy Violations in Agent Delegation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

AI agents use protocols such as Agent2Agent (A2A) to delegate tasks across independently built systems. Task policies govern which agents may use a tool or service and under what conditions. Runtime systems can check submitted operations against these policies. Before that check, the receiving agent chooses an operation using a handoff, the task and context sent with the delegation. If the handoff omits a required record, the agent may reject permitted work or attempt a forbidden operation. Agent evaluation must distinguish this information gap from a model error. To address this problem, we present PolicyFork, the first framework to verify whether agent handoffs determine policy decisions. PolicyFork adds the fewest available fields needed to determine each candidate action's decision and rechecks the repaired handoff. We also construct decision-sufficient execution state (DSES), a shortest fixed-length encoding that communicates these decisions to the recipient. We implement eight software-delivery policies from GitHub, GitLab, Amazon Web Services, Azure, and Google Cloud in local A2A workflows. Our release tests show that repaired handoffs let a rule-based receiver accept 28/28 permitted state–action combinations, up from 16/28, with zero false Permits. In paired trials, adding records raises correct Permits on allowed requests from 0/48 to 48/48 for each of four model configurations. PolicyFork separates missing information from model errors and exposes policy violations that task-success scores miss. Our data and code will be made available at https://anonymous.4open.science/r/PolicyFork-ICLR27/

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
agent-to-agent (A2A), agent security, Large Language Model, multi-agent security, delegation

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

This paper studies policy checks when one agent delegates a task to another. If the handoff misses a needed record, the recipient cannot decide. Two histories with the same input but opposite decisions form a delegation fork.

Contributions:

  1. A model-independent checker that finds forks.
  2. A minimum field repair and a short decision encoding (DSES).
  3. Experiments with eight vendor-based policies and four LLMs.
Critical strengths
  1. A clean diagnostic that does not depend on any model. A fork is a pair of histories that no recipient can tell apart. It separates handoff gaps from model errors.

  2. The checker is well validated. Mutation testing catches 50 of 53 modified policies. A separate predicate agrees with the monitor on 5,870 histories.

Critical weaknesses
  1. The formal core closely matches known results on observability and sensor selection in supervisory control (Lin and Wonham, 1988; Yoo and Lafortune, 2002). The paper does not discuss this link. The repair is shown only on two small cases that can be solved by hand.

  2. The model experiments are too small and too easy. The prompts are short JSON with at most two records and no distractors. Under Policy, permitted and forbidden requests share one input, so 0/48 only shows that models fail closed.

Questions for authors
  1. How does handoff sufficiency relate to observability?
  2. How do the results change if the model sees its own earlier actions?
  3. Can you test LLMs with free-text records, distractors, and the repaired handoff?
Reviewer_dg1COpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents PolicyFork, a framework for determining whether handoffs in delegated agent workflows contain sufficient information to resolve policy decisions before a recipient agent selects an action. The authors formalize a delegation fork as two policy-valid execution histories that produce the same complete recipient input but require opposite Permit/Deny decisions for the same candidate action. PolicyFork detects such conflicts, computes a minimum field repair by selecting the fewest available record fields needed to distinguish all conflicting histories, and rechecks the repaired handoff across all candidate actions. The paper further introduces decision-sufficient execution state (DSES), a shortest fixed-length encoding of the policy-decision profile over the candidate-action interface, together with a resource-side policy enforcement point that verifies signed DSES and the submitted request before execution. Experiments instantiate eight source-derived software-delivery policies from GitHub, GitLab, AWS, Azure, and Google Cloud in local A2A workflows. The evaluation measures handoff sufficiency before and after repair, compares model decisions with and without cross-agent execution records, studies model-chosen trajectories in which task success may coexist with policy violations, and evaluates DSES compilation and resource-side enforcement through exhaustive state enumeration and request replay.

Critical strengths

This paper has two clear strengths:

  1. This paper identifies a compelling failure mode in delegated agent systems: a recipient may make an incorrect policy decision because the handoff is information-theoretically insufficient, rather than because the model fails to use the information it receives. Separating missing information from model error is both conceptually important and practically useful, and the paper convincingly argues that task-success or policy-compliance scores alone cannot make this distinction.

  2. This paper gives an exact, model-independent criterion for handoff sufficiency and derives several results from a single decision-profile abstraction, including an impossibility result for insufficient inputs, minimum field repair, and a shortest fixed-length DSES encoding. The theory is especially effective because it maps directly onto the implemented checker, repair procedure, and enforcement mechanism rather than remaining purely abstract.

Critical weaknesses

Although the theoretical framework is elegant, two main weaknesses from experments currently limit my recommendation.

  1. All eight policies come from software delivery and share a similar structure, where a merge or deployment depends on checks, approvals, or version matching. It is unclear whether the conclusions hold in other domains, such as data access or policies with exceptions. The main 0/48 → 48/48 result compares deliberately incomplete inputs with clean records. The more challenging settings, answering Undetermined or querying missing records, have only eight trials per model. In addition, the evaluation does not include widely used frontier models such as GPT, Claude. Given that closed-source models can be readily evaluated through their APIs, including at least one such model would strengthen the results.

  2. Figure 2 reports only four invalid completions, all from the two GLM models, and the count is unchanged between Policy and History: each GLM model produces one invalid completion in each condition. Half of these violations occur when the relevant records have already been provided and are therefore more naturally attributed to model errors in using available information than to missing information. These examples establish that task success can hide a policy violation, but not that missing handoff information causes such violations or that handoff repair reduces them. The main observed effect of adding History is fewer incomplete trajectories and more valid completions, rather than fewer policy violations.

Questions for authors
  1. Can you provide results on policies outside software delivery, and on at least one widely used frontier model such as GPT, Claude, or Gemini?
  2. Could you report more trials for the Undetermined and record-query settings?
  3. Why does adding History leave the number of invalid completions unchanged in Figure 2? Do you claim that missing handoff information increases policy violations, or only that it makes the correct decision impossible to determine?
Reviewer_FCGeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

PolicyFork studies whether an agent’s handoff contains sufficient information to determine policy decisions for agent actions. It identifies conflicting execution histories, computes minimum field repairs, and constructs a compact decision-sufficient execution state. Experiments on software-delivery workflows evaluate handoff sufficiency, agent decisions, task completion, and resource-side enforcement.

Critical strengths
  • It's useful and important to distinguish between an agent’s reasoning error vs an information gap in its input. Model-independent counterexamples make this distinction testable, helping avoid misleading conclusions about policy-following ability.

  • The formulation and implementation are sound and reasonable. Checking complete recipient inputs and all candidate actions, distinguishing missing records from absent events, and validating against an independent evaluator make the framework credible within its stated scope.

Critical weaknesses
  • I feel the additional utility of minimum repair and DSES is not sufficiently established. The main LLM comparison supplies missing records rather than isolating the benefit of automatically selecting a minimum repair. Indexed verdict bitsets match DSES’s encoding optimum, and Online enforcement matches its replay results. Without demonstrated savings in communication, latency, or operational cost, the practical advantage of these central components remains unclear.

  • Gap between the contribution argument and the actual demonstration: the evaluation supports a narrower conclusion than the broader agent-delegation motivation suggests. The policies largely share approval/check prerequisites, and the central repair experiment has only 32 states and three actions. The 48 permitted judgments per model repeat eight distinct allowed inputs. Moreover, Figure 2’s valid-completion improvement, from 70/128 to 124/128, comes from fewer incomplete runs. Invalid completions remain at two per condition. That is, the value of supplying relevant records is demonstrated, but I woudl say that demonstration of evidence about discovering consequential information gaps in less curated workflows or reducing policy violations is limited.

Questions for authors
  • Which LLM results evaluate automatically selected minimum repairs rather than the full History condition? Is there a matched comparison with complete relevant records or on-demand retrieval that isolates the repair method’s benefit?
  • From the existing execution logs, how many Policy-condition interruptions were caused by output-budget exhaustion rather than decisions to stop because information was unavailable?
  • What is the intended operational advantage of DSES over indexed verdict bitsets or online policy queries, and what evidence supports that advantage beyond semantic bit counts?

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

Related papers

Discussion (0)

Sign in to comment.