Artificial Intelligence · arXiv:2610.04178 v1 · submitted 3 Oct 2026 · 40 references
SHarP: Saliency-based Pruning of Agent Harnesses
The paper says
“Agent harnesses are systems that coordinate model calls, tool use, and task execution to help large language models complete complex tasks.”
more of the abstract
To meet task requirements and address failures, these systems are often iteratively refined by amending and patching their instructions, tools, and workflows, continuously increasing harness complexity. It is therefore unclear whether some resulting harness modules are redundant, introducing substantial token overhead with little, if any, performance gain. Inspired by neural network pruning, in this paper, we study harness pruning as a means of striking a better balance between task performance and token cost. We propose SHarP (Saliency-based Harness Pruning), a simple yet effective pruning strategy based on the saliency of each harness module with respect to performance and efficiency. Specifically, we first identify tools, instructions, and supporting mechanisms as components that can be individually disabled. We then estimate the saliency of each module by ablating it and assessing its task performance and token cost relative to the full set of single-module ablations. Modules with the smallest contribution to performance or largest computational overhead are subsequently pruned. Our evaluation across various harnesses on held-out validation sets reveals a surprising finding: most harnesses that we studied are highly redundant and can maintain comparable performance and efficiency even after a substantial portion of their modules are pruned. Our pruning approach and empirical findings provide new perspectives on agent harness design and optimization.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 6 Oct 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. Solid, honest empirical paper on pruning agent harnesses; the token-cost savings are real, but the headline "comparable performance" and "improvement" claims rest on point estimates with no error bars. This line is wrong
4 major objections · 4 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 40 references say.
The paper says, at §4.2, Table 18
The review did not quote this passage.
Pith says: Objection 1 of 4 · major
The performance channel is reported only as point estimates with no confidence intervals or non-inferiority margins. The flagship result, OpenHands on GAIA improving from 29.44% to 35.00% (Table 18, [b116]), is based on 60 validation tasks; with three trials per task, the standard error of the aggregated pass rate is roughly 3.5 percentage points, so the reported gain is near 1.5 SE — indistinguishable from noise at conventional levels.
more of objection 1
The same issue affects the 'comparable performance' claim everywhere: e.g., PaperQA2's S15 drop from 61.40 to 56.14 (Table 17), and JIT S17 at 70.40 vs. full at 76.10 (Table 15), all sit within plausible noise bands. Because 'retains comparable performance' is the central, load-bearing claim of the paper, the manuscript should report per-configuration intervals (e.g., bootstrap or cluster-robust standard errors over tasks) and state a pre-specified non-inferiority margin, or explicitly downgrade the claim to 'no large degradation observed.' Without this, the abstract's claim that 'two harnesses even improve performance' ([b39]) is unsupported by the reported evidence.
The paper says, at §3.3, Eq. (5); Appendix A.3
The review did not quote this passage.
Pith says: Objection 2 of 4 · major
The protected module set is gated on p≤0.05 from one-sided paired t-tests on a single trial per module over n=30–50 development tasks. With binary success metrics and these sample sizes, the tests have very low power: a module whose removal genuinely costs a few points of pass rate will frequently fail to reach significance and be declared safe to prune.
more of objection 2
The paper's own development tables illustrate this — e.g., in Table 12, several OpenHands ablations show success-rate drops of 4–10 points (Search tool 44% vs. 62% median, Delegation 52%) yet only some reach p≤0.05. The measured validation performance is then the only remaining guard against harmful pruning, and, per the previous comment, that guard is too imprecise to certify retention. The authors should either use a less stringent protection criterion (e.g., effect-size or cost-based thresholds rather than p≤0.05), report the power analysis, or otherwise demonstrate that the protected set is not the sole mechanism carrying the safety of the procedure.
The paper says, at §3.2, Eq. (2); Appendix A.6
The review did not quote this passage.
Pith says: Objection 3 of 4 · major
The greedy per-module pruning order assumes that single-module saliency predicts multi-module outcomes, but the paper's own Appendix A.6 documents a concrete violation: the cost of PaperQA2's automatic evidence gathering 'depends on whether evidence retrieval is enabled,' so the efficiency saliency of one module is not invariant to which other modules are retained.
more of objection 3
The median-of-ablations reference in Eq. (2) inherits this fragility — the baseline itself shifts as modules are removed. The paper reports some evidence that the greedy order is not wildly wrong (qualitative trajectories), but it should either present a quantitative check of the additive assumption (e.g., comparing predicted vs. measured cost/performance for a few multi-module configurations) or qualify the pruning-order claim as heuristic. As written, the assumption that per-module rankings transfer to the pruned sequence is load-bearing for the efficiency-oriented results in Figures 2–3.
The paper says, at §4.3, Table 1; §4.1
The review did not quote this passage.
Pith says: Objection 4 of 4 · major
The pruning-strategy comparison in Table 1 uses one trial per validation task, in contrast to the default R=3 in Section 3.4, with no uncertainty reported on either side. The observed differences used to characterize the two strategies (e.g., efficiency-oriented retail S43 at 87.5% vs. all-pruned at 85.0% in Table 20, and performance-oriented DeepPlanning S'19 at 82.5% vs. full at 75.8%) are therefore subject to even larger sampling variation than the main results.
more of objection 4
At minimum, the paper should state the standard error for the one-trial validation estimates or justify why the strategy comparison is robust to the reduction in trials.
4 minor objections
The paper says, at Title/Abstract
Pith says: minor objection 1 of 4
The method is called 'SHarP' in the title and abstract of the arXiv listing but 'SHARP' throughout the body text; the notation should be made consistent.
The paper says, at §3.3, Eq. (5)
Pith says: minor objection 2 of 4
The α=0.05 threshold is applied per module with no multiple-comparison correction across m∈{18,…,48} modules; a brief justification or a sensitivity check on α would clarify how robust the protected-set composition is to the choice of threshold.
show 2 more
The paper says, at §4.1, Metrics
Pith says: minor objection 3 of 4
Table 7's validation sizes are small (20–60 tasks) and vary across settings; the paper should state per-setting standard errors or confidence intervals for the headline performance numbers in a single table, rather than requiring the reader to derive them.
The paper says, at Appendix A.5, Table 17
Pith says: minor objection 4 of 4
The all-pruned PaperQA2 configuration (21.93% at 30.6k tokens) is trivially cheap because the agent answers directly without tools; the caption or Section 4.2 should note that this baseline is not a meaningful 'pruned harness' but rather a no-tool control.
What would settle it
Compute a single paired per-task test per setting between the full harness and the headline pruned configuration on the held-out validation tasks (GAIA S3 vs S27; LIFE-harness retail S43 vs S48; JIT agentfold S20 vs S23; PaperQA2 S15 vs S18), using each task's three-trial mean as the unit and a 95% bootstrap CI on the mean difference. Declare 'comparable performance' only if the lower bound of the CI exceeds a pre-specified −5 percentage-point non-inferiority margin, and support any 'performance improves' claim only if the CI excludes zero. If, as the sample sizes suggest, the GAIA +5.6pt and the other differences have CIs spanning zero and dipping below −5pt, the redundancy finding's performance component is unverified and the abstract's claims must be re-scoped to the token-cost reduction alone.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 6 Oct 2026
- Reader holds up, with conditions
- Skeptic partly agrees
- Editor wrote the verdict above
- Referee major revision: 4 major objections, 4 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
This is a competent, honest applied paper. The genuinely new bit is applying ablation-based saliency pruning to heterogeneous agent-harness modules, with a two-axis saliency (performance and efficiency) and a protected set so critical modules survive. Prior work pruned agents and communication links or optimized harness configs, but not the fixed multi-component harness by contribution to both cost and performance. That framing is fair, and the evaluation is thorough: five settings, three trials on held-out validation, detailed module tables, and honest qualitative analysis showing pruning can increase cost (PaperQA2's completion tool, the evidence-gathering dependency in Appendix A.6). The token-cost side is the strongest part. Reductions like OpenHands 799k to 580k and JIT 329k to 197k are large and robust across trials, so the at-least-10% savings claim holds and is the practical win.
The soft spot is the performance-retention half of the redundancy claim, and it is load-bearing. On 30-60-task validation sets with three trials, a pass-rate standard error of 3-5 points makes "comparable" and even the GAIA 29.44 to 35.00 "improvement" statistically indistinguishable from noise. The paper is careful to say "observed" in Section 4.2, but the abstract and intro drop the hedge and assert comparable performance across all four harnesses. The stress-test note is right: with no error bars or non-inferiority margin, the performance half cannot be certified at these sample sizes. There is also a real circularity burden in the protected set: modules whose single-trial removal is significant at p<=0.05 are never pruned, so retention is partly engineered, and the low-power tests (n=30-50) can declare genuinely important modules safe to prune. The held-out validation numbers are then the only guard, and they are too imprecise to do the job. Appendix A.6's documented inter-module dependencies confirm that the evaluated multi-module configurations are the right object of inference, which is exactly where uncertainty quantification is missing.
None of this is fatal. The paper already acknowledges that pruning scope varies and that all-pruned configs can collapse (GAIA drops to 14.44%). The fixes are standard: add confidence intervals or a non-inferiority bound on the validation comparisons, run a sensitivity analysis on the protected-set threshold, and temper the redundancy claim to "the harnesses we studied." I would send this to a serious referee. With uncertainty quantification added, it is a genuinely useful contribution to agent efficiency; as-is, the token savings are real but the headline performance-retention claim is unproven.
The strongest objection
The central novelty is redundancy, which requires that pruned configurations actually retain performance; yet every validation comparison is a point estimate with no interval, and at these sample sizes the performance channel is unverifiable. On GAIA, 60 validation tasks × 3 trials give a pass-rate standard error near 3.5 points, so the showcase 29.44→35.00% (Table 18; [b116]) is about 1.5 SE — indistinguishable from zero, so nothing establishes that pruning improves, or even retains, performance. The abstract's 'two harnesses even improve performance' ([b39]) is unsupported on this evidence. Symmetrically, PaperQA2's observed S15 drop from 61.40 to 56.14 sits within noise, so the paper's honest observation that PaperQA2 trades cost for performance ([b129]) is itself unquantified, and 'comparable' is asserted without any non-inferiority margin. The token-cost side is solid: reductions like 799k→580k (OpenHands) and 329k→197k (JIT) are large and robust and do support the ≥10% cost claim. But the performance half of the redundancy claim — the load-bearing half — cannot be certified at n=30–60 tasks with 3 trials when a ±5pt degradation is undetectable. This is compounded by the protected set (Eq 5) being gated on p≤0.05 from single-trial, n=30–50 one-sided t-tests (Appendix A.3) whose low power means genuinely load-bearing modules can be declared safe to prune; the measured retention on the held-out validation set is then the only guard, and it is too imprecise to certify. Appendix A.6's documented dependencies (automatic evidence gathering's cost depending on whether evidence retrieval is enabled; OpenHands web-tool substitution) confirm that the evaluated multi-module configurations, not single ablations, are the right object of inference — which is exactly why the absence of uncertainty quantification on those configurations is the decisive gap.
I read the paper in good faith: the design is honest, the held-out evaluation is genuine (dev set disjoint from validation), single-module ablations are evaluated per task, and the token-cost reductions reported for pruned configurations are large, measured quantities whose magnitude dwarfs sampling noise. The method is cheap, the appendices (A.1–A.6) are unusually thorough, and Appendix A.6 deserves specific credit for documenting module interactions rather than hiding them. None of this is in dispute.
My concern is narrower and, I think, the single most load-bearing piece of the argument. The paper's headline ('most harnesses are substantially redundant') is only as strong as the performance-retention channel of its measurements, and that channel is presented without any uncertainty quantification. On validation sets of 30–60 tasks with three trials, pass-rate standard errors are on the order of 3–5 percentage points. A claim that a pruned harness 'maintains comparable performance' or 'improves task performance' is therefore not distinguishable, at the reported precision, from a claim that it lost several points. The GAIA improvement (29.44→35.00) is the paper's most striking retention datum and it is about 1.5 standard errors from zero. The 'comparable performance' framing throughout Section 4.2 is asserted, never tested against a margin. This matters more than the ordering concern the reader emphasized: the greedy per-module ordering determines which configurations are evaluated, but the evaluated configurations are directly measured on held-out tasks, so a wrong ordering would show up as a performance drop in the measurements — it cannot be the source of a spurious 'retention.' The source of a spurious retention is purely statistical: the measurements cannot resolve a real 5–10 point degradation. I therefore disagree slightly with the reader's emphasis: the low-power protected-set gating (p≤0.05, single trial, n=30–50) is a real compounding factor, but the decisive gap is the absence of intervals or tests on the validation comparisons themselves.
I am deliberately not manufacturing additional objections. The token-cost reductions, the Pareto-frontier analysis, and the qualitative trajectory evidence are sound and give the paper real empirical value regardless of the statistical gap. The verdict should remain CONDITIONAL: the paper is acceptable as a well-executed empirical study of cost reduction, but the redundancy/perfomance-retention claim requires paired per-task significance tests or non-inferiority bounds before it can be taken at face value. The reader already conditioned on this class of concern, so I do not move the verdict; I reinforce it and sharpen the test.
The referee's report
The paper's methodological contribution — a simple, evaluation-efficient pruning pipeline for heterogeneous agent-harness components — is clearly specified, and the appendix is unusually complete: candidate module tables with disabling semantics, full single-ablation saliency tables, per-configuration pruning logs, and token-accounting rules are all provided. If the redundancy finding holds, it would be a useful and somewhat surprising result for the agent-harness community. The token-cost reductions are large and robust (e.g., Table 18: 799k→580k on GAIA; Table 15: 329k→197k on DeepPlanning-Shopping) and do support the ≥10% cost-reduction claim. However, the load-bearing half of the claim — performance retention and even improvement — is certified only at the level of observed means on 20–60 validation tasks with three trials, at which the standard errors are several points, so 'comparable performance' and the flagship 'improvement' are statistically indistinguishable from no change.
Pith's reading of the paper (a machine's summary, not the authors')
This paper claims that agent harnesses — the layered sets of tools, instructions, and support mechanisms that orchestrate an LLM's model calls and tool use — accumulate removable redundancy, and that a simple saliency-based pruning scheme can exploit it. Borrowing from neural network pruning, the authors disable one module at a time on a small development task set, measure the shift in task score and token cost against the median of all single-module ablations, and rank modules by those two saliency signals. They report that across four harnesses and five settings, pruning away a large fraction of modules holds observed task performance roughly level while cutting token cost by at least 10%, and two harnesses even improve on both metrics. If the paper is right, harness redundancy is a cheaper design lever than adding capabilities, and systematic module auditing should accompany the usual practice of patching new tools and instructions onto a growing harness.
What carries the argument. The central mechanism is the two-axis saliency score built from single-module ablations. Each module is disabled once and the harness rerun; task score and token cost are recorded per task and compared with the per-task median over all single-module ablations, producing paired differences tested by one-sided t-tests whose p-values define 'performance saliency' and 'efficiency saliency.' The efficiency-oriented rule protects modules with performance p ≤ 0.05, then ranks the rest by efficiency p-value from high to low, yielding a sequence of pruned harnesses. The method's load-bearing step is the claim that this per-module greedy ranking, estimated from one trial per module on 30 to 50 development tasks, predicts what happens when many modules are disabled together.
Core claim. On the paper's own terms, the central claim is that most evaluated harnesses are substantially redundant, and that saliency scores from single-module ablations suffice to tell which modules can go. The authors test this with two one-sided paired t-tests per module — one on the change in task score, one on the change in token cost, each relative to the per-task median of all single-module ablations — and prune greedily: a protected set of performance-critical modules (performance p ≤ 0.05) is never touched, and the remaining modules are removed in order of worst efficiency saliency first. On validation sets disjoint from the development tasks, all four harnesses retain comparable performance with at least a 10% token-cost reduction, and OpenHands on GAIA and LIFE-harness on tau2-bench retail find pruned configurations that beat the full harness on both performance and cost. The paper also documents the other side: over-pruning hurts, with GAIA success falling from 35% to 14.44% when all 27 modules are removed, and removing termination or retrieval modules in PaperQA2 can raise token cost by prolonging execution or expanding processed text.
The premise it rests on. Saliency scores estimated from single-trial, single-module ablations on only 30 to 50 development tasks reliably rank every module's true contribution to performance and token cost, so the greedy per-module pruning order — with a protected set gated at p ≤ 0.05 from these low-power tests — predicts what happens when many modules are removed together.
If the paper is right (Pith's inferences, not the authors')
In all five settings a pruned configuration retains comparable observed performance with token cost reduced by at least 10%, so the measured harnesses carry removable redundancy.
Two settings (OpenHands on GAIA, LIFE-harness on tau2-bench retail) have pruned configurations that dominate the full harness on both performance and token cost.
Redundancy is uneven: PaperQA2 performs best at the full harness and loses accuracy when pruned, while LIFE-harness stays competitive even with all candidate modules disabled.
Over-pruning is a real failure mode: GAIA success drops from 35% to 14.44% at the all-pruned extreme, and removing the completion tool or evidence retrieval in PaperQA2 can raise token cost.
Efficiency-oriented and performance-oriented pruning reach different points on the performance–cost frontier, and the better strategy depends on the harness and domain.
Because saliency comes from a single trial per module on 30–50 development tasks, the ranking's stability is untested; re-estimating saliency on two disjoint development folds would show whether the protected set and pruning order are reproducible.
The greedy per-module order ignores second-order interactions that the paper itself documents (automatic evidence gathering's cost depends on whether evidence retrieval is enabled); a few pairwise ablations could test whether joint removals match single-module predictions.
The finding suggests harness-evolution pipelines could audit redundancy as a native step rather than after the fact, pruning before each new tool or instruction is added.
Because efficiency saliency can be negative — pruning a completion tool raises tokens, as observed in PaperQA2 — any automated pruning loop should protect modules whose removal increases cost even when it leaves performance unchanged.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Agent harnesses are systems that coordinate model calls, tool use, and task execution to help large language models complete complex tasks. To meet task requirements and address failures, these systems are often iteratively refined by amending and patching their instructions, tools, and workflows, continuously increasing harness complexity. It is therefore unclear whether some resulting harness modules are redundant, introducing substantial token overhead with little, if any, performance gain. Inspired by neural network pruning, in this paper, we study harness pruning as a means of striking a better balance between task performance and token cost. We propose SHarP (Saliency-based Harness Pruning), a simple yet effective pruning strategy based on the saliency of each harness module with respect to performance and efficiency. Specifically, we first identify tools, instructions, and supporting mechanisms as components that can be individually disabled. We then estimate the saliency of each module by ablating it and assessing its task performance and token cost relative to the full set of single-module ablations. Modules with the smallest contribution to performance or largest computational overhead are subsequently pruned. Our evaluation across various harnesses on held-out validation sets reveals a surprising finding: most harnesses that we studied are highly redundant and can maintain comparable performance and efficiency even after a substantial portion of their modules are pruned. Our pruning approach and empirical findings provide new perspectives on agent harness design and optimization.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-10-06
The authors' main claim, as the classifier picked it out 1.00
“Our evaluation across various harnesses on held-out validation sets reveals a surprising finding: most harnesses that we studied are highly redundant and can maintain comparable performance and efficiency even after a substantial portion of their modules are pruned.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.14 · human participants 0.03 · controlled comparison 0.71 · randomized assignment 0.02 · states quantitative result 0.09 · compares with prior results 0.23 · mentions code or data availability 0.03
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Agent harnesses are systems that coordinate model calls, tool use, and task execution to help large language models complete complex tasks. | background 1.00 | 0.04 | 0.01 | asserted 1.00 |
| To meet task requirements and address failures, these systems are often iteratively refined by amending and patching their instructions, tools, and workflows, continuously increasing harness complexity. | background 0.98 | 0.07 | 0.02 | asserted 0.87 |
| It is therefore unclear whether some resulting harness modules are redundant, introducing substantial token overhead with little, if any, performance gain. | gap 0.92 | 0.06 | 0.03 | hedged possibility 0.55 |
| Inspired by neural network pruning, in this paper, we study harness pruning as a means of striking a better balance between task performance and token cost. | contribution 0.92 | 0.30 | 0.04 | asserted 0.74 |
| We propose SHarP (Saliency-based Harness Pruning), a simple yet effective pruning strategy based on the saliency of each harness module with respect to performance and efficiency. | contribution 0.87 | 0.85 | 0.06 | asserted 1.00 |
| Specifically, we first identify tools, instructions, and supporting mechanisms as components that can be individually disabled. | method 0.99 | 0.67 | 0.02 | asserted 0.93 |
| We then estimate the saliency of each module by ablating it and assessing its task performance and token cost relative to the full set of single-module ablations. | method 1.00 | 0.94 | 0.03 | asserted 0.98 |
| Modules with the smallest contribution to performance or largest computational overhead are subsequently pruned. | method 1.00 | 0.79 | 0.06 | asserted 0.99 |
| Our evaluation across various harnesses on held-out validation sets reveals a surprising finding: most harnesses that we studied are highly redundant and can maintain comparable performance and efficiency even after a substantial portion of their modules are pruned. | result 1.00 | 0.97 | 0.84 | asserted 1.00 |
| Our pruning approach and empirical findings provide new perspectives on agent harness design and optimization. | interpretation 0.98 | 0.50 | 0.14 | asserted 0.93 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.99
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-06: 1 request, 7,540 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (5)
Figures are the authors', reproduced from the paper.
References (40)
- [1] Grace, Mikaela and Hadfield, Jeremy and Olivares, Rodrigo and De Jonghe, Jiri , title =. 2026 , month = jan, howpublished = on Pith
- [2] Trivedy, Vivek , title =. 2026 , month = mar, howpublished = on Pith
- [3] Meta-harness: End-to-end optimization of model harnesses , author=. arXiv preprint arXiv:2603.28052 , year= on PithPith's reviewarXiv
- [4] Agentic harness engineering: Observability-driven automatic evolution of coding-agent harnesses , author=. arXiv preprint arXiv:2604.25850 , year= on PithPith's reviewarXiv
- [5] Self-harness: Harnesses that improve themselves , author=. arXiv preprint arXiv:2606.09498 , year= on PithPith's reviewarXiv
- [6] Harbor: Automated harness optimization , author=. arXiv preprint arXiv:2604.20938 , year= on PithPith's reviewarXiv
- [7] Herding CATs: ALARA for agent harness engineering in portable composable multi-agent teams , author=. 2026 2nd International Conference on Human-AI Interaction and Experience Design (HAXD) , pages=. 2026 , organization= on Pith
- [8] Darts: Differentiable architecture search , author=. arXiv preprint arXiv:1806.09055 , year= on PithPith's reviewarXiv
- [9] Aviary: training language agents on challenging scientific tasks , author =. arXiv preprint arXiv:2412.21154 , year = on PithPith's reviewarXiv
- [10] Language agents achieve superhuman synthesis of scientific knowledge , author =. arXiv preprint arXiv:2409.13740 , year = on PithPith's reviewarXiv
- [11] PaperQA: Retrieval-Augmented Generative Agent for Scientific Research , author =. arXiv preprint arXiv:2312.07559 , year = on PithPith's reviewarXiv
- [12] Adapting the Interface, Not the Model: Runtime Harness Adaptation for Deterministic LLM Agents , author=. arXiv preprint arXiv:2605.22166 , year= on PithPith's reviewarXiv
- [13] JIT-Agent: Scaling Harness Intelligence via Just-in-Time Harness Evolution , author=. 2026 , eprint= on Pith
- [14] The OpenHands Software Agent SDK: A Composable and Extensible Foundation for Production Agents , author=. 2026 , eprint= on Pith
- [15] GAIA: a benchmark for General AI Assistants , author=. 2023 , eprint= on Pith
- [16] ^2 -Bench: Evaluating Conversational Agents in a Dual-Control Environment , author=. 2025 , eprint= on Pith
- [17] Zhang, Yinger and Jiang, Shutong and Li, Renhao and Tu, Jianhong and Su, Yang and Deng, Lianghao and Guo, Xudong and Lv, ChenXu and Lin, Junyang. D eep P lanning: Benchmarking Long-Horizon Agentic Planning with Verifiable Constraints. Proceedings of the 64th Annual Meeting of ... on Pithdoi:10.18653/v1/2026.acl-long.335
- [18] Han, Song and Pool, Jeff and Tran, John and Dally, William J. , title =. Proceedings of the 29th International Conference on Neural Information Processing Systems - Volume 1 , pages =. 2015 , publisher = on Pith
- [19] Michel, Paul and Levy, Omer and Neubig, Graham , booktitle =. Are Sixteen Heads Really Better than One? , url = on Pith
- [20] Jonathan Frankle and Michael Carbin , title =. 7th International Conference on Learning Representations,. 2019 , url = on Pith
show 20 more
- [21] Externalization in llm agents: A unified review of memory, skills, protocols and harness engineering , author=. arXiv preprint arXiv:2604.08224 , year= on PithPith's reviewarXiv
- [22] Natural-language agent harnesses , author=. arXiv preprint arXiv:2603.25723 , year= on PithPith's reviewarXiv
- [23] Code as agent harness , author=. arXiv preprint arXiv:2605.18747 , year= on PithPith's reviewarXiv
- [24] Aflow: Automating agentic workflow generation , author=. International Conference on Learning Representations , volume= on Pith
- [25] Automated design of agentic systems , author=. International Conference on Learning Representations , volume= on Pith
- [26] n.d. , howpublished = on Pith
- [27] Rcr-router: Efficient role-aware context routing for multi-agent llm systems with structured memory , author=. arXiv preprint arXiv:2508.04903 , year= on PithPith's reviewarXiv
- [28] Resum: Unlocking long-horizon search intelligence via context summarization , author=. arXiv preprint arXiv:2509.13313 , year= on PitharXiv
- [29] Agentfold: Long-horizon web agents with proactive context management , author=. arXiv preprint arXiv:2510.24699 , year= on PitharXiv
- [30] Snip: Single-shot network pruning based on connection sensitivity , author=. arXiv preprint arXiv:1810.02340 , year= on PithPith's reviewarXiv
- [31] Learning structured sparsity in deep neural networks , author=. Advances in neural information processing systems , volume= on Pith
- [32] Pruning filters for efficient convnets , author=. arXiv preprint arXiv:1608.08710 , year= on PithPith's reviewarXiv
- [33] Learning efficient convolutional networks through network slimming , author=. 2017 IEEE international conference on computer vision (ICCV) , pages=. 2017 , organization= on Pith
- [34] Cut the crap: An economical communication pipeline for llm-based multi-agent systems , author=. International Conference on Learning Representations , volume= on Pith
- [35] Agentdropout: Dynamic agent elimination for token-efficient and high-performance llm-based multi-agent collaboration , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages= on Pith
- [36] Adaptive graph pruning for multi-agent communication , author=. ECAI 2025: 28th European Conference on Artificial Intelligence 25-30 October 2025, Bologna, Italy-Including 14th Conference on Prestigious Applications of Intelligent Systems (PAIS 2025) Proceedings , pages=. 2025... on Pith
- [37] AgentSlimming: Towards Efficient and Cost-Aware Multi-Agent Systems , author=. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages= on Pith
- [38] The complexity trap: Simple observation masking is as efficient as LLM summarization for agent context management , author=. arXiv preprint arXiv:2508.21433 , year= on PitharXiv
- [39] Beyond Prompts: Measuring and Optimizing LLM Tool-Agent Harnesses , author=. arXiv preprint arXiv:2609.05736 , year= on PithPith's reviewarXiv
- [40] Efficient Memory Management for Large Language Model Serving with PagedAttention , author=. Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles , year= on Pith
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment