This Week in Agentic AI: Harnesses Become the Primary Optimization Surface

Across capability, efficiency, and safety, the strongest gains came from redesigning the systems around models rather than scaling the models themselves.

Weekly Research Digest28 August 2026

Agent performance increasingly depends on the harness: generated and evolved runtimes produced large gains without model retraining, while structured retrieval and context pruning reduced cost without sacrificing accuracy. The safety papers reinforce the same architectural lesson—guarding, evaluating, and debugging individual decisions is more reliable than trusting final answers or aggregate scores.

LV-NUS Lab

Generating Agent Harnesses on Demand Improves Models Across Benchmarks — A model trained to generate task-specific memory, planning, action, and tool orchestration improved multiple model families and matched established runtimes such as OpenCode and Claude Code in controlled evaluations, making harness synthesis a credible alternative to hand-built agent stacks.

Appier AI Research + National Taiwan University

Joint training helps small language models create and use tools — Jointly training a 4B Qwen3 model to create and invoke tools yielded 79.8% macro-average accuracy across 13 held-out procedural tasks and transferred to unseen tabular and visual QA, showing that tool construction can be learned as a general reasoning skill.

Eindhoven University of Technology + Dana-Farber Cancer Institute

Medical AI models often give right answers with disconnected reasoning — Across 30 clinically motivated perturbations and four medical QA benchmarks, answers often remained unchanged when displayed rationales were meaningfully altered, demonstrating that correct outputs do not establish chain-of-thought faithfulness.

Shanghai Artificial Intelligence Laboratory + Beihang University + Fudan University + Renmin University of China + KAUST

Step-level checks curb unsafe agent actions with little utility loss — StepGuard evaluates risky trajectory steps before tool execution and reduced mean attack success on AgentDojo and AgentDyn by 77.3%, with only a 2.8-point reduction in mean utility.

Independent / Undisclosed

Persistent corpus maps help agents find evidence within tight budgets — AtlasNav uses a persistent map of corpus structure to counter evidence blindness, reaching 92.05% strict accuracy on BrowseComp-Plus while cutting recorded online inference cost by 30.21% versus the previous dynamic-workspace state of the art.

National University of Singapore + TCG CREST + IIT Kharagpur + Singapore Institute of Technology + IIT Bombay

Bounded citation-graph search improves recall while cutting scholarly search costs — Crase searches a predefined citation neighborhood with an explicit stopping boundary, delivering up to 3× the recall@50 of proprietary deep-research agents at roughly one-third of their cost on a 500,000-paper corpus.

Stanford University + University of California, Santa Barbara + Prime Intellect + University of Washington

Sliding reasoning windows make long test-time scaling 3x faster — Prefix Sliding retains the instruction prefix and recent reasoning while discarding stale intermediate tokens, producing up to 3× speedups without performance loss and enabling trained models to exceed 100,000-token reasoning traces.

Fudan University + Tencent

Search agents improve longer when their critics evolve alongside them — CAFE jointly evolves search behavior and corrective feedback in a shared-parameter model, continuing to improve across in- and out-of-domain evaluations after agent-only or critic-only training had plateaued.

Tel Aviv University + AWS Agentic AI

Adaptive trace graphs improve failure attribution in multi-agent systems — Adaptive Influence Graphs convert multi-agent traces into traversable causal structures and achieved state-of-the-art Who&When results, giving diagnostic agents a more precise substrate for locating which component failed and when.

Bar-Ilan University + UNC Chapel Hill + University of Texas at Austin

In AI-Q, most report errors originate in the orchestrator — Invocation-level attribution across three deep-research systems found that orchestrators introduced 84.7% of final-report errors; targeted fixes based on those diagnostics improved citation recall by 5% without reducing output quality.

ServiceNow + Mila + Université de Montréal

Evolved agent harnesses lift enterprise performance without retraining models — Failure-guided search over harness designs improved three enterprise benchmarks by 20–35 percentage points, with gains transferring to held-out tasks and across GPT and Qwen model families.

AWS

Switching Models Mid-Task Imposes a Cost and Quality Penalty — In long-running coding tasks, handing a weaker model's full trajectory to a stronger model recovered less than half the quality gap while adding substantial cost, exposing a handoff tax that naive model-routing policies miss.

ATMRI + Nanyang Technological University + Centre of AI Research + VinUniversity + IIT Mandi

Semantic scores overstate language-model safety in air traffic control — An aviation-standard benchmark developed with 40 controllers found that semantic metrics systematically obscure the unequal consequences of operational errors; risk-aware fine-tuning narrowed but did not eliminate the discrepancy.

Alibaba Group + Harbin Institute of Technology

Contrastive branch training improves credit assignment for tool-using language models — CBPO separates where alternative continuations are sampled from how token-level credit is assigned, achieving the highest macro-average accuracy among evaluated policy-optimization and branch-based methods across ten reasoning and search benchmarks at two model scales.

Zhejiang University + Zhongguancun Academy + University of Science and Technology of China + National University of Singapore

Reinforcement learning strengthens policy invocation for agent safety judgments — RePolicy trains a safeguard to identify and invoke the policy relevant to an entire execution trajectory, maintaining strong safety detection across six benchmarks even when policy contexts changed.

Creative AI & Agentic Generation

Self-evolving loop synthesizes high-quality multimodal training data — vivo AI Lab's VISA uses verifier signals and target-model failure profiles to iteratively generate harder multimodal instruction data, improving MM-IFEval while preserving general capability across seven additional benchmarks.

Analysis

The harness is not plumbing; it is a learnable optimization surface. LV-NUS generates task-specific harnesses at runtime, while ServiceNow, Mila, and Université de Montréal evolve harnesses from observed failure patterns and obtain 20–35-point enterprise gains without retraining the underlying model. This extends the now 30-paper agent-harness thread from manual runtime design toward automated systems search; teams should version, evaluate, and optimize harnesses as independently as model weights.

More context is not the answer; better context topology is. AtlasNav's persistent corpus map reaches 92.05% strict accuracy while lowering online cost, and Stanford-led Prefix Sliding discards most old reasoning tokens for up to a 3× speedup without accuracy loss. The shared result is that agents need durable structure and selective access, not indiscriminate context accumulation; teams should separate persistent indexes, immutable instructions, and short-lived working traces.

Open-ended search is not inherently agentic; bounded structure can be both better and cheaper. Independent researchers behind AtlasNav prevent evidence blindness with a persistent corpus representation, while the NUS-led Crase system constrains scholarly exploration to a citation neighborhood and reports up to 3× higher recall@50 at roughly one-third the cost. Retrieval agents should begin with an explicit search graph, budget, and stopping condition before adding autonomous query loops.

Agent safety is a trajectory-control problem, not a final-answer classifier. StepGuard from Shanghai AI Laboratory and collaborators blocks risky actions before execution, Zhejiang University's RePolicy selects the applicable policy from full trajectories, and the Eindhoven–Dana-Farber audit shows that even correct medical answers may have disconnected rationales. Safety teams should enforce policy at action boundaries and treat displayed reasoning as untrusted telemetry rather than proof of compliance.

Safety metrics must price consequences, not linguistic similarity. The ATMRI and NTU air-traffic study finds that semantic scores hide the unequal operational costs of different errors, while StepGuard reports attack reduction alongside utility loss rather than collapsing both into one accuracy number. For consequential agents, evaluations should attach domain-specific severity to each action and report security, utility, and catastrophic-risk metrics separately.

The orchestrator is both the leverage point and the failure concentration point. The AI-Q analysis from Bar-Ilan, UNC, and UT Austin attributes 84.7% of report errors to orchestrators, while AWS finds that mid-task model switching recovers less than half the expected quality gap because trajectory transfer itself imposes a handoff tax. Teams should test routing, state serialization, and recovery logic as first-class components rather than assuming stronger downstream models can repair weak orchestration.

Better agent learning comes from assigning supervision to decisions, not just outcomes. Appier and NTU jointly train tool creation and use, Fudan and Tencent co-evolve a search agent with its critic, and Alibaba with Harbin Institute of Technology separates branch sampling from token-level credit assignment. Training pipelines should expose tool design, branch choice, critique, and policy invocation as distinct learning targets instead of relying on a single terminal reward.

Multi-agent debugging needs causal structure, not larger log files. Tel Aviv University and AWS turn traces into adaptive influence graphs for state-of-the-art failure attribution, while the AI-Q study shows that invocation-level provenance can identify targeted orchestrator fixes that improve citation recall. Production systems should record typed dependencies among agents, messages, evidence, and tool calls so diagnostics can traverse causes rather than search flat transcripts.