acceptodds
Under review as a conference paper at ICLR 2027

STATE DYNAMICS OF AGENT HARNESSES:WHAT STATES MATTER FOR PRESERVATION?

Abstract

Agent harnesses shape long-horizon agent behavior by controlling base model and tool calls, feedback integration, and persistent workspace state. As a result, the same base model can exhibit different interaction dynamics and task outcomes across Harnesses. Yet existing Harness designs are still shaped largely by empirical rules, leaving the propagation of local changes over a finite interaction window, the conditions that a state representation must preserve for later operations, and the formation of a readable task object without a common dynamical account. In this work, we develop a conditional state dynamics framework that formalizes finite response, conditional attractor structure, operation-aware state fidelity, and task readout. Analyses across models, harnesses, and tasks provide empirical support for the relationships among spectral structure, interaction persistence, successor prediction, and code readout. Building on the task-object criterion, we propose BasinGuard, an online plugin that runs at Harness protocol events, continues the host protocol while the object is still forming, and invokes the specified readout once the object is determined. Across the evaluated model–Harness conditions, BasinGuard raises the normalized task score by 17.15% on average on code workspaces and improves official pass rates on GAIA2.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Agent harness, LLM Agents, State Abstraction

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_qFRTOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The contributions of this paper lie along two axes:

  1. Theoretical analysis of harness dynamics: The paper models harness interactions as absorbing Markov chains and studies protocol-state dynamics, state aggregation, and predictive fidelity.

  2. BasinGuard, an online harness plugin: The authors propose a plugin that uses a counter-based logistic gate and heuristic context management to improve agent performance. BasinGuard consistently improves success rates across three harnesses and two models, increasing GitTaskBench success rates from 9.33% to 22.00% without increasing runtime.

Critical strengths

S1. BasinGuard consistently improves success rates across multiple harness–model combinations without increasing execution time, demonstrating its practical effectiveness and generalizability.

Critical weaknesses

W1. The theoretical results are disconnected from the proposed method. Theorems 1 and 2 essentially restate standard identities and the strong lumpability condition, without providing a theoretical foundation for BasinGuard, which instead relies on a counter-based logistic gate and heuristic context management.

W2. The empirical validation lacks comparisons with existing techniques and component-level ablations.

Questions for authors

Q1. The analysis assumes that the protocol state is Markov (control-sufficient). Appendix D.2 shows that adding the previous state st−1s_{t-1} greatly improves next-event prediction (Brier 0.4687 → 0.1887), which suggests that the current state sts_t alone is not Markov. How do the authors reconcile this with the first-order kernel used to estimate ρ^\hat\rho, α\alpha, and the residence scales in Section 2.1? Also, can the authors clarify what the spectral and attractor analysis adds beyond a 2 * 2 substochastic matrix? Similarly, strong lumpability (Theorem 2) holds or fails by construction in Section 2.2 rather than being tested on the estimated kernels.

Q2. BasinGuard is compared only against the native harness path; there are no component ablations (M/O/K), no predictor-free control (e.g., random-time or fixed-interval application of the same operations), and no comparison with existing context-management methods such as observation masking, LLM summarization (e.g., ACON), or the host’s own compaction, even though the anonymous code defines arms such as M, O, K, MOK, and O_RANDTIME. Can the authors report these comparisons, and clarify how the shared checkpoint x0x^0 is chosen (task start or mid-trajectory)?

Q3. How often did the gate actually fire (i.e., how often qθ≤q⋆q_\theta \le q^\star among otherwise eligible boundaries), and which operations were applied? If it fires at nearly every boundary, the predictor adds little beyond a fixed schedule. Please also report how qθq_\theta and the threshold and scheduling parameters (q⋆,a⋆,c⋆,j⋆q^\star, a^\star, c^\star, j^\star) were obtained, whether they were tuned on the evaluation tasks, and their values, which are missing from both the paper and the anonymous code.

Additional Comments

A1. Figures 2 and 4 are computed from synthetic constructions (the discrete double-well kernel of Appendix C.2 and the three-variable family of Appendix C.4), not from the interaction logs, but neither the captions nor the main text (L124, L201-207) says so, and Appendix C.2 and C.4 are never cited near figures. Please state explicitly in both captions that these are synthetic illustrations, cite the corresponding appendix sections, and give the numerical kernel of the C.4 family, since Figure 4 cannot currently be reproduced from the text.

A2. Details of figures and experiment setup can be improved. (e.g., number in abstract is only Qwen3.8 results, not the average.)

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

The paper models agent harnesses using conditional state dynamics, connecting finite-response kernels, state abstraction, and task readout. It proposes BasinGuard, a plugin that manages history and tool obligations before reading out a task object, and reports gains on GitTaskBench and GAIA2.

Critical strengths
  1. The paper emphasizes a useful distinction between producing program text and completing the workspace facts required by task evaluation. The figures help illustrate why preserving operation-relevant state can matter.

  2. BasinGuard reports positive results across two models and three harnesses, including an overall +10.87-point normalized GitTaskBench score change and improvements in official GAIA2 pass rates.

Critical weaknesses
  1. The connection between the theory and the implemented controller is not established. Theorems 1-2 concern transition perturbations and strong lumpability, whereas BasinGuard uses a logistic predictor and rules for history release, re-anchoring, and tool recovery without using the spectral or fidelity quantities. The correlations and recoding experiments do not show which new design decision follows from the theory. Extensive notation makes this gap harder to assess; a concrete trace connecting formal states to actual controller actions is needed.

  2. The evaluation does not isolate BasinGuard's contribution. Table 2 compares only the native harness with the full plugin, without simple history-management/completion baselines or component ablations. Appendix E.1 reports duration ratios, but leaves token/tool budgets, predictor training, and the meaning of the "frozen public-state cascade" unclear. These omissions make both the source of the gains and the realism of the paired comparison difficult to judge.

Questions for authors
  1. Which BasinGuard action or threshold is actually derived from the spectral/lumpability analysis? What does that analysis predict beyond a simple rule to preserve pending obligations and produce required artifacts?

  2. Does the public-state cascade execute fresh model/tool calls after intervention, replay observations, or fix later continuations? Please provide the task counts, predictor training/labels, and token/tool budgets for both paths.

  3. Are there existing comparisons with simple truncation, deduplication, tool-retry, or completion checks? Please also clarify the independent unit of the paired sign test and give GAIA2 discordant counts or confidence intervals.

Reviewer_goA5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents a conditional state dynamics framework for understanding and improving long-horizon LLM agent harnesses. It investigates how local harness modifications propagate through interaction trajectories, which state distinctions must be preserved for subsequent operations, and when program and workspace states jointly constitute an evaluable task object. The authors formulate agent-harness interactions as locally absorbing Markov processes and characterize finite-horizon perturbation responses through conditional attractors and spectral properties. They further introduce operation-aware state fidelity, using strong lumpability conditions to characterize when compressed state representations preserve future transition behavior. A Brier-score decomposition quantifies the predictive information lost through state aggregation. These theoretical components are connected to task evaluation through an explicit readout map that incorporates program outputs, workspace artifacts, and execution feedback. Building on this formulation, the paper proposes BasinGuard, an online harness plugin that monitors public protocol states and applies structured continuation operations, including bounded history release, task re-anchoring, and pending-tool recovery. The controller continues execution while the task object remains unreadable and invokes the prescribed readout once an evaluable object has formed.

Critical strengths

S1. The paper addresses a meaningful challenge in long-horizon agent systems by studying how harness-level state transitions, information preservation, and task evaluation interact.

S2. The emphasis on preserving state distinctions required by future operations is particularly interesting. It provides a principled perspective on history compression and context management beyond generic similarity-based or recency-based preservation.

S3: BasinGuard demonstrates positive gains across multiple models and harness architectures. The paired evaluation protocol and improvements in both normalized task scores and official task success rates support the practical relevance of the proposed runtime interventions.

Critical weaknesses

W1. The most significant concern is that BasinGuard does not explicitly use the central theoretical quantities introduced in the paper, such as spectral residence, conditional attractor distributions, or operation-aware predictive costs. Its practical interventions appear largely driven by protocol-level heuristics and a logistic continuation predictor. Consequently, it remains unclear how the theoretical analysis directly informs the algorithmic design or explains the reported performance gains.

W2. The experiments primarily compare the complete BasinGuard system against native harnesses. Without isolating the contributions of history release, tool recovery, task re-anchoring, continuation prediction, and readout gating, it is difficult to determine which components drive the improvements.

W3. The empirical analysis relies on a small number of protocol states that may not capture the semantic complexity of real agent execution. The required Markov and control-sufficiency assumptions are not directly established. Furthermore, improvements in one-step successor prediction do not necessarily demonstrate preservation of long-horizon task outcomes.

W4. I think the proposed readable-state boundary characterizes whether an evaluable object exists, but not whether additional execution could improve its correctness. The paper should investigate readable-but-incorrect states and clarify when immediate readout is preferable to continued refinement.

Questions for authors

Q1. How does the theoretical framework actually inform BasinGuard's design?

Q2. The experiments mainly compare the complete system against native harnesses. Could the authors provide component-wise ablations and comparisons against simpler baselines, such as periodic context compaction or rule-based tool recovery, under comparable execution budgets?

Please see the above weaknesses w1-w4 for more details.

Reviewer_ZPf1OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a conditional state-dynamics framework that models agent–harness interaction as an absorbing Markov chain over protocol states, with theorems on how local rule changes propagate and when a compressed state representation preserves what later operations need. Building on it, the authors introduce BasinGuard, a plugin that hooks into an existing harness (Pi, Aider, DeepSeek Harness) and uses a logistic gate over protocol counters to trim or de-duplicate history, recover pending tool calls, or trigger task evaluation. With Qwen3.8-27B and Muse-Glimmer-30B, BasinGuard reports a +10.87 normalized-score gain on GitTaskBench (+17.15 for Qwen3.8) and small gains in GAIA2 pass rate.

Critical strengths

BasinGuard is evaluated as a drop-in plugin across three different host harnesses and two models with paired runs, and on GitTaskBench the absolute success rate rises from 9.33% to 22.00% (Appendix E.1).

Critical weaknesses
  • BasinGuard is essentially a context-management/compaction plugin, but it is compared only against the host harness without it. It is possible to adapt plugins from existing harnesses like OpenHands or Codex to make it clearer whether BasinGuard improves over standard compaction.
  • None of BasinGuard's components (the logistic gate and threshold, history release, re-anchoring, tool recovery, early readout) are ablated, and the paper offers no analysis of which component drives the gain
  • Results use a custom normalized score rather than GitTaskBench’s standard task pass rate, the number of evaluated tasks is not stated, and results aren't compared against published numbers.
  • Several gains have a standard deviation of exactly 0.00 across three runs (Appendix E.1) consistent with a single-task difference.
  • Only two open models are used, with thinking mode disabled
  • The dense notation also makes the paper hard to follow.
Questions for authors
  1. How many GitTaskBench tasks were evaluated, and what are the absolute counts and standard task pass rates behind each reported difference? Why are several standard deviations exactly zero?
  2. How does BasinGuard compare with an existing compaction method (e.g., OpenHands’ condenser, Codex CLI’s compaction, or ACON) plugged into the same harnesses?
  3. Can you ablate each component and report how often each operation fires?
  4. Why these two models with thinking mode disabled? Do the gains hold for a model with published GitTaskBench results?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.