acceptodds
Under review as a conference paper at ICLR 2027

Audit Horizons: How Long Can Stateful Agents Act Before Review?

Abstract

How many consecutive changes should a coding agent make before a trusted review? We introduce the audit horizon: the longest sequence of autonomous actions a fixed agent stack may take before trusted review, under a guarantee on regression risk. We then propose an audit-horizon qualification method: design data choose how deep and how many independent trajectories to run, and fresh trajectories determine the horizon through a valid risk-control test. We prove that one-shot evaluation cannot identify risk that accumulates along a trajectory. Certifying a nonzero audit horizon with high probability therefore requires persistent evidence. However, when earlier observations follow the same distribution regardless of the planned evaluation length, extending a run beyond the target horizon adds cost without strengthening its direct risk bound. On chained upgrades of Flask, Jinja2, and PyJWT with DeepSeek V4 Flash 0731, short-prefix qualification reaches the same audit horizon as full-length evaluation while reducing measured model cost by 53.6% and summed agent-environment time by 42.5%. Across a three-model policy comparison, sharing evaluation depth across packages achieves the highest observed horizon-cost score. Overall, we advocate that future safety metrics should reflect long-term behavior instead of only one-shot workloads.

Primary area:
datasets and benchmarks
Keywords:
LLM agents, coding agents, agent evaluation, long-horizon evaluation, stateful agents, AI safety, statistical risk control, evaluation efficiency

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_WEasOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces the audit horizon, defined as the maximum number of consecutive actions a coding agent may take autonomously before requiring trusted review, subject to a bound on cumulative regression risk. It proposes a statistical qualification procedure that uses persistent agent trajectories to select evaluation depth and replication, then applies fresh trajectories to certify the horizon. The method is evaluated on chained upgrade tasks from SWE-Chain involving Flask, Jinja2, and PyJWT.

Critical strengths
  1. This paper introduces the concept of an “audit horizon,” linking the number of consecutive actions taken by a coding agent to cumulative regression risk, and proposes a corresponding statistical qualification method.
  2. The experiments show that, in the evaluated setting, short-prefix evaluation achieves the same horizon as full-length evaluation at lower cost.
Critical weaknesses
  1. The introduction is too brief, and Figure 1, the only figure before the experiments, is not explained clearly enough, making the paper difficult to follow.
  2. The experimental evaluation is limited in scope and focuses primarily on SWE-Chain.
Questions for authors
  1. Figure 1 is not mentioned anywhere in the main text. Could the figure be explained more clearly?
2 comments

We thank the reviewer for the constructive feedback for our presentation. We agree that the introduction and Figure 1 should better explain the motivation and evaluation workflow. We will expand the introduction with a concrete example of a coding agent performing consecutive repository upgrades, illustrating why independently resetting the environment after each action cannot reveal regression risk accumulated across inherited repository states. We will also explicitly introduce and explain Figure 1 in the main text. The figure contrasts three evaluation strategies: independent clean‑reset trials, full persistent trajectories, and shorter persistent prefixes. The key distinction is that persistent prefixes preserve state within each trajectory while allowing more independent roots to be evaluated under a fixed resource budget. Regarding experimental scope, we agree that broader evaluation would strengthen the generality of the conclusions.

I thank the authors for their response; however, I still find that my concerns have not been resolved. For instance, the introduction remains too short, and no new experiments have been provided. Therefore, I have decided to maintain my previous score.

Reviewer_b1muOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces the audit horizon, the longest block of stateful agent actions allowed before trusted review under a cumulative regression-risk constraint. Design data select trajectory depth and replication, while fresh qualification trajectories determine the returned horizon through fixed-sample or sequential tests. On chained upgrades of three Python packages, the paper reports matched H2 horizons at lower short-prefix cost and the highest horizon-cost score for shared depth.

Critical strengths
  • The audit-horizon formulation is practically meaningful.
  • The separation of design and qualification data is statistically principled.
  • The persistent-evidence analysis provides a useful evaluation insight.
  • The presentation and resource accounting are clear and transparent.
Critical weaknesses
  • The central risk guarantee applies neither to repeated deployment nor to the admission-coupled evaluation process.
  • The policy and efficiency claims rely on retrospective archive replay, an inactive adaptive policy, and a single utility/seed.
  • The empirical scope and regression-only risk definition do not support the broader AI safety framing.
Questions for authors
  1. The deployment motivation exceeds the formal contract. Eq.(2) guarantees one action block from a prespecified initial-state law, while Section 8.1 states that review need not restore that law. What guarantee applies to the second and later blocks in the repeated review workflow motivating the paper?

  2. Upper-coding does not by itself resolve admission-induced intervention. Unknown outcomes are conservatively coded as failures, but denied requests can change which actions occur and can depend on shared spending history. Is admission independent of latent regression outcomes, and do the conclusions persist with prospectively uncontaminated data or a dependence-robust analysis?

  3. The adaptive mechanism is not empirically exercised. Section 7.2 reports that adaptive depth retains its initial depth on every observed path and is therefore identical to fixed depth. Please provide a setting in which adaptation changes a decision, or narrow the claimed contribution to fixed-depth selection.

  4. The policy savings are retrospective rather than realized. Section 6.3 reads policy-requested prefixes from a common depth-six archive, whose physical collection costs 18.163–19.547B qualification tokens and 9.941–11.121B validation tokens (Table 10). How would the policies compare when executed prospectively with their actual stopping and shared-budget dynamics?

  5. The shared-depth ranking is insufficiently robust. The objective uses a setting-specific H−λCH-\lambda C with λ=.02/C2\lambda=.02/C_2, one outer seed, and accounting intervals rather than sampling uncertainty. Please report sensitivity to λ\lambda, repeated seeds, and uncertainty in both horizons and costs.

  6. The theoretical novelty remains unclear. Theorem 5.1 is an identification characterization, Theorem 5.2 uses the first unsafe coordinate, Theorem 5.3 applies a standard data-processing lower bound, and Proposition 5.4 follows from prefix consistency. What new statistical or algorithmic principle is contributed beyond ordered risk testing and time-uniform sequential inference [1,2]?

  7. The sequential guarantee appears inapplicable to the collected archive. Section 4.2 and Appendix B.3 require prespecified conditional laws and conditional independence within a balanced round, yet Appendix D.7 reports admission failures in 155/540 qualification roots and 275/432 validation roots, and Section 8.1 states that the shared mechanism can couple trajectories. Please identify the result that covers this process or restrict the empirical guarantee accordingly.

  8. The operational meaning of "safe before review" is unclear. The experiments use α=.40\alpha=.40, and the event is regression under specified external tests; Section 8.1 notes that this excludes upgrade failure and untested harms. Why is this risk level and event definition appropriate for the claimed review decision rather than only for benchmark non-regression?


[1] Angelopoulos, A. N., Bates, S., Candès, E. J., Jordan, M. I., and Lei, L. (2021). Learn then Test: Calibrating Predictive Algorithms to Achieve Risk Control. arXiv:2110.01052.
[2] Howard, S. R., Ramdas, A., McAuliffe, J., and Sekhon, J. (2021). Time-uniform, nonparametric, nonasymptotic confidence sequences. The Annals of Statistics, 49(2), 1055–1080.

1 comment

We thank the reviewer for the detailed assessment. We agree that several distinctions between the theoretical contract and its empirical interpretation need to be made more explicit. We address the concerns in four groups.

  1. Scope of the guarantee (Q1, Q8).

Eq. (2) certifies a single autonomous action block from a prespecified initial‑state distribution. It of course does not imply an unconditional guarantee for subsequent blocks after review; Section 8.1 notes that review need not restore that distribution, and we will further clarify the single‑block constraint in the abstract and conclusion in the revision. Similarly, α=0.40α = 0.40 is an illustrative benchmark operating point for measured regression risk; we do not propose it as a production safety threshold, and we will clarify it in the revision.

  1. Admission failures and empirical validity (Q2, Q7).

We agree that upper‑coding unresolved outcomes does not by itself address admission‑ induced intervention or cross‑trajectory dependence. We would like to note that this only affects the immediate results from the experiments and not the theory. On whether admission is independent of latent regression outcomes: Yes, and we therefore 1. restrict the three‑model results to a description of the recorded experiment; 2. point to the direct resource comparison (Table 2), whose cost cap was scoped to a single transition and which returns and independently validates H=2H = 2 without shared admission; 3. planning to rerun the three‑model qualification and validation stages report the results in the revision.

  1. Policy comparison and efficiency (Q3–Q5).

The reviewer is correct that adaptive depth did not change its initial decision on any observed trajectory. Thus, the experiments do not demonstrate an empirical advantage from adaptation. We will distinguish the general adaptive‑depth formulation from the fixed‑depth benefits actually observed.

We do, however, maintain that archive replay is a valid way to evaluate depth policies. Each policy’s depth decision is recorded before its prefix environment is exposed, and the agent does not in fact know which depth it is at. The first h steps of a depth‑six root therefore have the same law as a root run to depth h (Proposition 5.4). Hence, the replayed prefixes are valid samples for evaluating each policy. Of course, replay omits some noise of the actual cost caused by re‑running the policy under a different seed.

  1. Theoretical and methodological novelty (Q6).

We do not claim new concentration inequalities, and we admit that the primitives we use are established. We would, however, not describe the results as standard conse‑ quences of well known textbook facts: Theorem 5.1 and Theorem 5.3 concern a decision object that, to our knowledge, has not been formalized before. Our main contribution is methodological: we defined audit horizon as a concept, and giving an evaluation procedure so that this horizon can be certified at low cost.

Reviewer_893uOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper asks how many consecutive actions a stateful agent can take before review while keeping cumulative regression risk below a chosen limit. It estimates this audit horizon from trajectory prefixes and evaluates the approach on chained package upgrades.

Critical strengths

The paper turns a practical question about when to review an agent's work into a clear statistical problem. The proofs appear sound under the stated assumptions, and the direct comparison shows that shorter evaluation runs can reduce cost.

Critical weaknesses

I am not convinced that the reported audit horizons are ready to use for execution without the budget interruptions seen in the experiments. Your analysis requires independent trajectories from a fixed process, but shared-budget denials can change what the agent does and introduce dependence between runs. These denials affect 30 of the 120 qualification trajectories used by the fixed/adaptive policies, even within the returned horizons (Appendix D.7). You acknowledge the problem in Section 8.1, but I still do not see evidence that these horizons remain valid without those interruptions.

Questions for authors

Do these audit horizons remain valid without the budget interruptions? What evidence supports that?

1 comment

We thank the reviewer for the careful reading and for the assessment of the theory. We agree with the core of the concern and would like to clarify below: The guarantee and its assumptions. The guarantee holds for fresh roots drawn independently from a fixed process law, seperate from experiment results. We agree that our experiments does not fully satisfy the assumptions to some extent. Seperate results without shared admission. The direct resource comparison (Table 2) did not use the shared mechanism. Its model‑cost cap was for a single transition, hence not affected by the problem stated. In that study, both the 100‑root prefix design and the 60‑root full‑length design return H=2H = 2, and the independent 200‑root validation sample gives U2=0.3449<0.40.U2 = 0.3449 < 0.40. To answer the question directly for the three‑model setting, we are rerunning the qualification and validation stages with abundant budget. We will re‑ port the resulting horizons and validation outcomes in the revision whatever they are, including any horizon that fails to requalify, and will adjust the corresponding claims accordingly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.