acceptodds
Under review as a conference paper at ICLR 2027

From Evidence to Effect: Authority Semantics and Runtime Infrastructure for Stateful Agents

Abstract

Stateful agents reuse artifacts after producing executions and permissions change. We formalize authority-sufficient observations and durable effects bound to execution and material identities. WTB implements this interface through runtime adapters, shared evidence, and transactional publication/recovery. Six study families separate the mechanism from its integration. Raw and typed evidence both solve 32/32 authority cases, with model-dependent planning effects. Fixed-intent enforcement blocks six unsafe proposals and executes 12 eligible authorized intents. Complete controls match WTB's capability. Paid integration yields 176/210 accepted benchmark-source stages, including 19/30 publication stages, recovery on 8/8 primary SWE repositories, and the most complete continuous trajectories on each of three source tasks. The findings connect authority information, effect admission, and infrastructure reuse in stateful agents.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
LLM Agents, Agentic Systems, Multi-Agent Systems, Agent Security, Authorization, Runtime Enforcement, Provenance, AI Governance

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

This paper introduces WTB, a framework for connecting execution evidence, authorization, and durable effects in stateful agents. It formalizes authority-sufficient observations and execution-bound publication, with runtime adapters and transactional recovery mechanisms. Experiments investigate information disclosure, authority enforcement, runtime responsibilities, and integration across coding tasks. Results show that WTB preserves authorization constraints and achieves higher integration success than the evaluated baseline packages under fixed budgets.

Critical strengths
  1. The paper addresses an important problem in stateful agent systems by connecting execution provenance, authorization, and durable effect enforcement through a unified framework.
  2. The experiments cover multiple aspects of the proposed system, including controlled authority tests, runtime ablations, and coding integration. WTB demonstrates strong authorization correctness and promising integration improvements.
Critical weaknesses
  1. Unclear attribution of integration gains. Complete baseline implementations achieve the same authority-contract coverage as WTB, while the coding comparisons use packages with different prebuilt capabilities. This makes it difficult to separate WTB's architectural advantages from its more directly reusable functionality.
  2. Limited independent validation. Despite numerous experimental runs, the coding evaluations rely on only three benchmark-source tasks and eight primary SWE repositories. Moreover, the adjusted statistical comparisons remain inconclusive, limiting confidence in generalization.
Questions for authors
  1. Can the authors clarify how much of WTB's integration advantage comes from its architecture rather than differences in the functionality supplied to each baseline?
  2. How robust are the observed integration improvements across additional independent tasks, repositories, and model configurations?
Reviewer_JjdFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies authorization for stateful agents that reuse artifacts across executions and changing permissions. It proposes WTB, a runtime framework that binds producing attempts, contributor identities, material versions, and approvals to durable effects. The framework checks authority when an effect is committed and supports recovery when an external operation’s outcome is uncertain. The evaluation is done with separation on access to authority information, its presentation to a model, runtime enforcement, and the effort required for coding agents to integrate and maintain the framework.

Critical strengths

S1. The paper addresses an important practical problem for multi-agent systems. Agents may reuse outputs after the original conditions or permissions have changed. The proposed framework provides a concrete way to manage these situations, which could help make longer agent workflows more reliable.

S2. The empirical evaluation considers several aspects of practical use. The experiments examine agent decisions, execution safety, and the effort needed to integrate and maintain the framework. The authors also report cases where the method does not outperform the alternatives, making the assessment reasonably balanced.

Critical weaknesses

W1. The main contribution to AI-agent research could be explained more clearly. Much of the paper concerns authority infrastructure design. While this is relevant, I find it difficult to identify the main new insight about agent behavior or the specific advantage over existing approaches to reliable agent execution. A clearer connection between the proposed design and its benefits for AI agents would strengthen the paper.

W2. The generalizalbility of the empirical advantages remains uncertain. The coding experiments mainly use one model configuration and a limited set of tasks. The results are encouraging, but it is unclear whether the same advantages would hold for other models and more diverse agent workflows.

Questions for authors

Q1. Could you explain the main contribution from an AI-agent perspective and clarify what benefit it provides beyond existing reliability tools?

Q2. Could you provide some additional evidence that the integration and maintenance advantages could be transferred to other coding models or a broader range of tasks?

Reviewer_79TuOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper seems to focus on authorization and integrity in agentic AI systems. The authors propose an architecture to preserve these properties. The contribution seems to be a mix of prof of concept evaluation and benchmark evaluation.

Critical strengths

None

Critical weaknesses

The paper is really hard to read and I am not sure what is the nature of the contribution.

The readability can be improve by gently introducing the concepts (e.g., workspace/material space, scoped authority, publication ledger) and mapping them to real-world system.

The problem under study is fundamentally a security problem, so for future revision, I would suggest authors to spend some time on giving the big picture (e.g., addressing certificate revocation in distributed systems), explain the threat model and specificity of agentic AI systems. This will also help in identifying system level baselines, which are currently missing.

Questions for authors

Cf. weaknesses

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.