News without the noise
Harnesses, orchestration, context engineering
61 articles · page 1 of 2
Li et al. designed AgentZip to reduce the memory consumed when agent tasks spawn many related sandbox sessions. Across LLM training and inference workloads, the system reduced sandbox-owned memory by up to 8.7-fold while limiting the slowdown from aggressive compression to 1.40-fold.
Yu, Fang, and Chen propose Agent-Integrated Software, a pattern combining a conventional application core, direct user interaction, and a built-in agent. Their transition-system framework connects user-facing task semantics to application execution, while interaction contracts and continuous assurance specify when delegated actions and outcomes remain valid.
Gao et al. developed VikingRAG, a retrieval-augmented generation system that navigates both the meaning and structure of documents while retrieving additional evidence only when needed. Across real-world datasets, the authors report token consumption of 11.6%–51.9% of state-of-the-art baselines, falling to 5.1%–32.5% when retrieval traces are reused and multi-round search is invoked selectively.
Wang et al. introduce SWRouter, a method for routing each turn of a multi-turn conversation to a suitable large language model. Across dialogue benchmarks, the authors report higher evaluation accuracy than both the best individual model and a strong conversation-context routing baseline.
Amirmoshiri, Sahneh and Jangjou built an agentic platform that transforms heterogeneous Chemistry, Manufacturing and Controls documents into a queryable, provenance-linked knowledge graph. On 505 questions drawn from 38 reports for one Sanofi small-molecule program, the lexical retrieval layer performed strongly on direct questions but weakened on comparative and corpus-wide queries.
The authors developed Ecdysis, a framework for improving the runtime code and instructions that govern an LLM agent’s execution. Rather than revising a harness after each failed task, it analyzes batches of failures for recurring patterns and uses multiple diagnostic roles to propose refinements. In the reported experiments, this reduced training time by up to 1.84× while increasing reasoning accuracy by 18.56% over existing harness-evolution methods.
Lu et al. frame agent-skill optimization as a budget-constrained search over a changing pool of candidates. Across six benchmarks and three target models, their method delivered the highest average performance among the tested approaches while requiring 55–58% less optimization cost than SkillOpt.
Marron and Barr propose Agentic Just-In-Time Software Construction, in which an application includes an AI agent that observes runtime behavior and modifies software logic or interfaces in response. Their demonstration presents trace-driven human–AI co-construction as an alternative to delivering software solely as a fixed, prebuilt artifact.
Chen et al. introduce an interface that lets vision-language models make fine-grained physical decisions using semantic action units, which deterministic interpreters translate into commands for particular robots. The authors report that this approach transfers across tasks, environments, and robot embodiments while outperforming representative agent-based and vision-language-action baselines.
Ma and Kordjamshidi introduce CLAMP, a decoding framework that converts visual scene evidence and a symbolic action model into constraints on an embodied planner. Across three benchmarks, the authors report better object grounding and safety, with remaining errors concentrated in perception and incorrectly specified constraints.
Schall et al. introduce Q2D-Web, a benchmark designed to reflect the first retrieval stage of agentic retrieval-augmented generation systems. Across lexical, dense, and late-interaction retrievers, model rankings remain broadly stable under different relevance-labeling methods but vary substantially by topic, language, and query type.
Li et al. introduce SkillAdam, a method for iteratively improving the natural-language instructions and procedural knowledge used by frozen language-model agents. Across seven short- and long-horizon benchmarks, the authors report more stable optimization, higher task performance, and fewer, less costly revision rounds than prior self-evolution methods.
Ben Amor et al. built RefactorPlatform, an open-source harness for comparing repository-scale refactoring agents under controlled conditions. Experiments across four model families show that AST-aware retrieval—organizing code by its syntactic structure—substantially improves success, while their multi-agent configuration trails a lean retrieval-augmented single agent.
Wang and colleagues developed TROVE, an orchestration method that learns reusable skills and outcome-dependent transitions from evaluated workflow traces. Rather than committing to a fixed plan or replanning from scratch, TROVE validates the route after each top-level skill and selectively edits only the continuation invalidated by new evidence.
Shi and colleagues built CUA-Universe, an environment-to-data pipeline that converts desktop software into reproducible settings where agents can act through both graphical interfaces and command lines. Training a 9B model on its verified trajectories substantially improved reported task success while reducing interaction steps and token use across three evaluations.
Wang et al. propose TAHI (test-time adaptation through human-agent interaction), which leverages a user's cross-session interaction history to update agent context and weights while maintaining an evolving rubric that captures personal evaluation criteria. Applied to writing and visual creation across 30 individuals, the adapted agents outperformed non-adapted baselines on individual tasks, and part of the personalization gains transferred to other users.
The authors built a multi-agent harness for long-running bioinformatics analyses in which plans, code, outputs, and supporting evidence persist throughout the workflow. On three benchmark tracks, Bioinfoysis substantially improved accuracy across multiple underlying language models, including reaching 82.4% on BixBench.
Fan et al. developed an environment-evolution method that modifies terminal tasks along three difficulty-related directions and schedules successive generations during training. Applied to two Qwen models using long-horizon reinforcement learning, the method improved Terminal-Bench 2.1 performance by 14.4 and 18.0 percentage points.
The authors built an agent that distills operational know-how from repositories and papers, then retrieves those skills while carrying out machine-learning research tasks. With the GPT-5.5 backbone, harness, and execution budget held constant, adding skills raised benchmark scores by 9.2% to 134.3% relative to the same agent without them.
The authors built CivBench, an open-source Civilization VI environment in which language-model agents operate through 76 Model Context Protocol tools over episodes exceeding 300 turns. Tests across four model families revealed persistent gaps in proactive state monitoring and plan execution, although the small sample does not support reliable model rankings.
Zhou et al. analyze deep-research agent trajectories and identify premature commitment as a recurring failure mode: early tool calls steer later searches toward the same potentially misleading path. Their HypoSearch method uses lightweight hypotheses to organize parallel exploration, outperforming single-trajectory and standard parallel-search baselines across four benchmarks and three backbone models.
Wei introduces a three-level routing framework that converts prior agent trajectories into reusable skills instead of rerunning an LLM reasoning loop for every request. Across security-monitoring and ToolBench evaluations, the authors report token reductions of 62.3% and 76.3%, respectively.
The authors benchmarked six LLMs on building agent harnesses from minimal seeds and then revising those harnesses using downstream feedback. Across 2,207 held-out instances, generated systems sometimes matched human-engineered references, but results varied sharply by domain, execution cost, and the model running the harness.
Pakhomov and Nijkamp built a live trace system that incrementally converts an agent’s event history into typed state and separate views for agents and human observers. In synthetic evaluations, these views made monitoring more accurate and economical, while deterministic state aggregation prevented failures on long sequential-dependency tasks.
Yan and colleagues developed Harness-of-Harness, a supervisory framework designed to help coding agents improve a software project across repeated development loops without human intervention. Across three benchmarks and three harness–model pairings, the authors report a 52.25% average relative gain over standalone harnesses, alongside a multi-day demonstration lasting more than 70 iterations.
The authors developed a training-free agent that searches long videos differently depending on whether a question requires localized evidence, broad coverage, or comparison between hypotheses. Across four benchmarks, this policy-steered approach generally outperformed a uniform-retrieval baseline, including gains of 2.7 accuracy points on Video-MME-v2 and 6.9 points on LongVideoBench’s long-video subset.
Wu and colleagues introduce ASPIRE, an interactive benchmark in which agents must interpret a natural-language capability goal, choose how to improve, and evaluate their own progress. Across six goals and 520 hidden expert-authored items, agents completed self-modification loops but produced sparse, unstable weight gains and harnesses that remained behind a human-engineered reference.
The authors introduce a benchmark in which a controller model repeatedly reviews structured progress summaries and directs a separate coding worker, including deciding what to verify and when to stop. The best controller achieved only a 24.69% strict success rate on full tasks, while a lower-cost partial-task evaluation closely reproduced the full evaluation’s model ranking.
The authors evaluated model-generated self-modifications on 600 unseen, one-shot tasks and found 197 mutations that improved capability but failed recoverability checks. Controlled interventions showed that most failures could be repaired only by combining precise state references with a more expressive language for specifying recovery actions.
The authors analyzed 16,000 real-world sessions, finding that 75.9% involved multiple turns, then developed a framework that simulates users with different personas and interaction behaviors. Training Qwen3.5-27B on 6,300 generated records raised its benchmark composite score by 4.1%, while reducing the turns and tool calls needed to complete tasks.
Ye and Lin instrumented an agentic datasheet-extraction benchmark to record every tool call, then built rules for attributing failures and detecting suspicious runs from tool use alone. Their detector flagged all 50 deliberately planted tool-withholding faults while raising no flags across 207 clean, fidelity-passing extractions, although its ability to catch other failure types remains untested.
Pan et al. introduce ContextPilot, a framework for controlling the growing working context of agents performing long-horizon tasks. On long-context question answering and deep-search benchmarks, the authors report better task performance than existing baselines while retaining a more compact context.
Song, Kwak and Chang present String, an open-source runtime that describes application views, typed actions, navigation and credentials in a single String-Flavored Markdown document. Across 87 tasks and six language models, String matched the aggregate success of curated agent skills while completed episodes used 33.5% fewer tokens.
Zhu and Shi present DSA, an LLM-agent orchestration framework for producing stock-research reports from heterogeneous evidence across multiple markets. Their evaluation establishes that a frozen implementation conforms to selected software contracts, but does not test whether its reports improve forecasts, investment decisions, or returns.
The authors identify “instruction privilege escalation,” in which an agent harness places untrusted content into a more privileged part of a model’s context. Using multi-agent attack mechanisms, they achieved all 13 tested security objectives on every evaluated harness under unrestricted execution, and on all three supporting automatic permission review.
Xu et al. developed HarnessLens, a framework that automatically modifies agent instructions, tools, and runtime components while limiting the cost of testing each change. Across three harnesses and four benchmarks, it improved average held-out performance by 7.6–13.6% while using substantially less evaluation budget than competing methods.
The authors train a model to generate an agent’s harness—the memory, planning, action, and tool-orchestration machinery surrounding an underlying language model—for each task at execution time. They report that these generated harnesses consistently improve multiple model families and can match established runtimes such as OpenCode and Claude Code in controlled evaluations.
Huang et al. compare general-purpose agents with specialized agents for identifying the root causes of system incidents, then introduce an external harness that learns from previous diagnoses. Across two public benchmarks and an industrial deployment, OpsHarness reaches 59.0% top-1 accuracy, a 63.4% improvement over an unharnessed general agent.
Li, Abdelmoniem, and Wang introduce ProgRouter, an online routing framework that adapts model selection as a multi-agent task unfolds rather than making a single decision at the outset. Across coding, mathematics, and retrieval-augmented question answering benchmarks, the authors report lower operating costs than key baselines while retaining strong task performance.
Liang et al. modify speculative decoding so its drafting and verification models can see different versions of an agent’s accumulated context. Across four agentic capabilities and two end-to-end benchmarks, the method retained about 90% of full-context accuracy on average while reporting substantial throughput and compute gains on isolated text tasks.
Xu and Chen isolate context construction as an optimization problem for long-video understanding. Their outer-loop system proposes, executes, and evaluates context-building programs around a frozen vision-language model, finding improvements from both a uniform-sampling start and a stronger hand-crafted baseline.
Burke presents a working literate-programming environment designed for collaboration between human programmers and language-model agents. It co-locates code with relevant prose and structured data while exposing symbol-like name relationships that agents can search and navigate.
Guo et al. introduce a structured tool-use policy that separates reconstructing task state, deciding whether to act, specifying an admissible action, and producing the final output. Across multi-turn, multi-tool, and incomplete-information evaluations, the authors report more reliable task completion than direct function calling and ReAct, especially for smaller models and tasks dependent on earlier context.
The authors distinguish two sources of parallelism in long reasoning traces: decomposing a problem into independent subtasks and exploring multiple speculative trials at once. They restructure sequential traces with a formal grammar, train models to balance accuracy and latency, and report about 1.7× average wall-clock acceleration on AIME24 and AIME25 while retaining competitive accuracy.
Guo, Zheng, and Cao study “Evidence Blindness,” in which an agent can technically access relevant evidence but fails to surface, open, or inspect it within a limited interaction budget. Their AtlasNav framework uses a persistent Corpus Atlas and reaches 92.05% strict accuracy on BrowseComp-Plus, with 30.21% lower recorded online inference cost than the previous dynamic-workspace state of the art.
Chen and colleagues introduce PeakBench, a benchmark of executable multi-tool workflows annotated with dependencies and measured resource demands. Their evaluation shows that agents capable of sound logical planning do not necessarily schedule tools safely or efficiently, while access to resource information can reduce avoidable capacity overflows.
Kim et al. built Simthesizer, a framework in which a harnessed coding agent modifies a shared simulator of LLM serving workflows rather than constructing a separate simulator for each new mechanism. The authors report 2.51% average throughput error against a vLLM-based system and simulation speeds up to 284.96 times those of competing simulators.
Hazra et al. introduce Crase, a scholarly search system that explores a predefined citation neighborhood rather than letting an agent search iteratively without a fixed stopping rule. Across two benchmarks on a 500,000-paper arXiv corpus, the authors report up to three times the recall@50 of proprietary-model deep research agents at roughly one-third of their cost.