Topic
agents
ToolPro: Executable Tool Programs Cut Web Service Latency by 53% for LLM Agents
A new system called ToolPro replaces static API endpoints with executable tool programs for LLM-based agents, achieving up to 53.4% lower latency and 96.1% less traffic on multi-step workflows. It uses effect-aware replay for exactly-once execution and a profile-driven policy to decide when program execution outperforms stepwise calling.
The Autonomy Tax: Defense Training Breaks LLM Agents
A new research paper reveals that defense training designed to protect LLM agents from prompt injection attacks paradoxically destroys their ability to perform multi-step tasks, causing 99% timeout rates and worse security than undefended baselines. The study identifies three systematic biases and attributes them to shortcut learning.
LedgerAgent: A New Method for Policy-Adherent Tool-Calling AI Agents in Customer Service
Researchers introduce LedgerAgent, an inference-time method that maintains observed task states in a separate ledger and checks policy constraints before tool calls, improving pass^k metrics across four customer-service domains. The approach addresses common failure modes where agents use stale or incorrect information or violate domain policies.
AutoPass: Evidence-Guided LLM Agents Achieve Compiler Speedups of 1.117x on ARM64
Researchers present AutoPass, a multi-agent LLM framework that uses compiler and runtime evidence to guide compiler optimization decisions. Without training, it outperforms expert-tuned heuristics and classical autotuning, achieving geometric-mean speedups of 1.043x on x86-64 and 1.117x on ARM64 over LLVM -O3.
Automated Skill Mining from GUI Trajectories Shows Readability but Limited Policy Improvement
A team of researchers proposed a three-stage pipeline to automatically generate skill libraries for computer-using agents by mining interaction trajectories. While the mined clusters achieved high purity (five of eight clusters ≥0.95) against expert labels, downstream policy improvements were marginal: GRPO raised skill-step accuracy only from 18.5% to 20.5% and failed to outperform trivial frequency priors on key source-domain metrics. The study is presented as a diagnostic, underscoring that readability does not guarantee transfer.
Beyond Static Leaderboards: Predictive Validity for Evaluating LLM Agents in Enterprise AI
A new paper on arXiv proposes replacing static aggregate-score leaderboards with predictive validity—correlation between in-sample and out-of-sample rank—for evaluating LLM agents. The authors argue that current benchmarks underspecify deployed-agent evaluation, based on fourteen parallel implementation studies and seven prior agent benchmarks. They introduce a twelve-tier measurement apparatus and falsifiable out-of-distribution criteria.
SkillsVote Framework Governs Agent Skills Lifecycle from Collection to Evolution
SkillsVote is a lifecycle-governance framework for Agent Skills that addresses redundancy, quality, and environment sensitivity in open skill ecosystems. It profiles a million-scale open source corpus, decomposes trajectories into skill-linked subtasks, and admits only successful reusable discoveries to evidence-gated updates. Experiments on Terminal-Bench 2.0 and SWE-Bench Pro show performance improvements on challenging agentic coding benchmarks.
TERMS-Bench Diagnoses LLM Negotiation Agents Beyond Deal Rate for Enterprise Procurement
A new benchmark called TERMS-Bench goes beyond deal rate to diagnose why LLM negotiation agents fail, evaluating 13 frontier models on surplus extraction, cue use, belief calibration, and compliance. For enterprise procurement and trade, this offers actionable insights into AI agent weaknesses.
AutoDojo: Adaptive Attacks Expose Superficial Defenses and Structural Limits in LLM Agents
The AutoDojo framework adaptively optimizes indirect prompt injections against LLM agent defenses, revealing that many current defenses are superficial. Against a filter that reduces static attack success rate to 0%, AutoDojo recovers 28% overall and 64% on action-open tasks due to a structural limitation where injections can pose as ordinary data.
LectūraAgents Multi-Agent Framework Promises Adaptive Personalized AI-Assisted Learning
Researchers propose LectūraAgents, a multi-agent framework for adaptive personalized AI-assisted learning. It uses a hierarchical architecture with a ProfessorAgent leading specialized agents to generate and deliver tailored lecture content with embodied teaching actions. The system was validated on diverse courses and showed gains in content quality and personalization.
User as Code: Executable Memory Paradigm Boosts AI Agent Personalization with Deterministic Execution
A new arXiv preprint introduces User as Code (UaC), a paradigm where AI agent user memory is stored as executable Python code rather than unstructured text. On benchmarks, UaC achieves 78.8% recall on LOCOMO and 99% accuracy on aggregate questions, while enabling unsolicited safety alerts that retrieval-based systems cannot provide.
XFlow: A New Programming System for Reliable Multi-Agent Workflows Addresses Prompt–Harness Boundary
Researchers present XFlow, an executable protocol programming system designed to improve reliability in LLM-based multi-agent workflows. By introducing the XPF protocol language and lifecycle-governed symbols, XFlow makes constraints and process requirements explicit and enforceable, addressing the underspecified prompt–harness boundary that limits current systems.
CoffeeBench: New Benchmark Evaluates LLM Agents in Multi-Agent Economic Simulations
Researchers introduce CoffeeBench, a benchmark for evaluating LLM agents in a long-horizon multi-agent economy. The 90-day simulation features farmers, roasters, and retailers, with models controlling one roaster. All models outperformed a passive baseline, but Claude Haiku 4.5 showed an idle-drift failure mode.
Agentomics Framework Introduces Shapley Value-Based Pricing for AI Agents in Human-AI Workflows
A new paper from arXiv introduces Agentomics, a workflow-based framework that applies coalition game theory and Shapley value to value, attribute, and price AI agents in human-AI teams. The framework models workflows as heterogeneous agent configurations, addressing complementarities and bottlenecks, and uses a security-operations case study to demonstrate productivity gains and reliability losses.
S1-DeepResearch: New AI Agent Combines Search and Synthesis for Long-Horizon Research Tasks
Researchers introduce S1-DeepResearch, a unified framework for training deep research agents that combine closed-ended QA with open-ended exploration. The 32B-parameter model achieves state-of-the-art among open-source models across 20 benchmarks spanning reasoning, instruction following, report generation, file understanding, and skills usage.
LLM Agents Look at Correct Tools but Still Pick Wrong, Research Reveals Readout as Failure Point
Research by Shiyang Chen reveals that LLM agents mis-call tools not because they fail to see the right tool, but because the decision readout fails. The model attends to the correct tool 80% of the time, yet picks wrong. Readout-side interventions recover 59-91% of failures, while input-side fixes recover ≤23%.
RetailBench Benchmark Tests LLM Agents on Long-Horizon Retail Decisions
Researchers introduced RetailBench, a simulation benchmark for evaluating LLM agents in single-store supermarket management over 180 days. Tests on seven models showed only a subset completed the full horizon, and even the best fell far behind an oracle policy due to incomplete evidence acquisition and lack of consistent strategy.
New Framework Automates Skill Construction for Agentic Large Language Models
A new framework called Collective Skill Tree Search (CSTS) automatically constructs reusable skills for large language model (LLM) agents. It uses two iterative phases—collective generation and collective assessment—to build a diverse, generalizable tree of skills that enhances agentic capabilities in planning, tool use, and environment interaction.
Risk-Aware LLM Agents for Geospatial Data Retrieval: New Framework Passes Adversarial Tests
Researchers present a risk-aware LLM agent framework for retrieving remote sensing data from cloud-based geospatial catalogues using natural language queries. The system integrates Guardrail, General-QA, and Recommender-Analyst agents to convert user intent into structured API calls. Preliminary adversarial evaluation shows prompt-level safety instructions improve robustness, though rare high-impact failures persist.
ToolMenuBench: New Benchmark Evaluates Tool-Menu Filtering for Reliable and Efficient LLM Agents
ToolMenuBench, a new benchmark from researchers, evaluates how tool-menu filtering strategies affect LLM agent reliability and efficiency. In tests across seven model backends, causal minimal tool filtering improved task success from 32.1% to 85.7% while reducing token usage by roughly 98%.
ACC Method Compiles Agent Trajectories to Enhance Long-Context Reasoning in LLMs
Researchers propose Agent Context Compilation (ACC), which converts agent trajectories from search, software engineering, and database tasks into long-context question-answer pairs. Training Qwen3-30B-A3B with ACC achieves 68.3 on MRCR and 77.5 on GraphWalks, matching a model 8x larger, while preserving general capabilities.
E-mem: Multi-Agent Framework for Episodic Memory Reconstruction Boosts LLM Reasoning Efficiency by 70%
Researchers propose E-mem, a multi-agent framework that reconstructs episodic context for LLM agent memory, inspired by biological engrams. It uses a hierarchical architecture with assistant agents maintaining uncompressed contexts and a master agent orchestrating planning, achieving 54% F1 on the LoCoMo benchmark, surpassing the state-of-the-art GAM by 7.75% with over 70% token cost reduction.
TrustedARI: A New Trust-Native Infrastructure Secures Agentic AI Routing for Enterprise Deployments
TrustedARI, presented by a research team on arXiv, is the first trust-native agentic routing infrastructure for agentic AI. It addresses fundamental trust risks in agent routing, offering a 39.34% reduction in handshake overhead and verifiable billing with 28.20x faster proof generation, all without modifying service providers.
APEX Adaptive Principle Extraction Framework Enables Multi-Dimensional Self-Evolution for Production AI Agents
Researchers propose APEX (Adaptive Principle EXtraction), a three-layer self-evolution framework that simultaneously improves an AI agent's prompt harness, behavioural principles, and workflow topology. Tested on the production-grade Joe AI agent built on NVIDIA Nemotron, APEX achieved a 90% improvement in Health Score over baseline, distilling six novel reusable principles and selecting a research-first workflow scoring 0.900 (+20%). The framework outperforms single-axis harness optimisation and requires only 4 LLM calls (~270 seconds).
Technology Pine Labs Launches India’s First Agent-to-Agent UPI Payment Protocol for AI Deals
Pine Labs has launched the Pine Labs Payment Protocol (P3P), India's first agent-to-agent payment capability built on UPI mandate architecture. The system allows users to authorize an AI agent to independently execute transactions based on predefined conditions, removing the need for human authentication at the point of transaction.
Technology Half of workers worry AI will still take their job as agent usage soars 90% in a year
New data from GMB Union reveals nearly half of UK workers worry AI will take their job, amid a 90% year-over-year increase in AI agent usage reported by Stack Overflow. Despite growing adoption, most organisations still require human oversight for autonomous agents, and concerns about accuracy and security persist.
Technology How AI Agents Can Protect EV Charging Infrastructure from Cyberattacks
Researchers from the NICS lab at the University of Malaga have developed a system using multiple AI agents to protect electric vehicle charging infrastructure from cyberattacks. The agents collaborate using a consensus mechanism based on opinion dynamics to provide a comprehensive view of the network's security state. The proposal aims to detect anomalies early and prevent attacks ranging from energy theft to larger grid disruptions.
Logistics AI Agents Transforming Freight Operations
AI is revolutionizing freight operations by automating tasks and enhancing decision-making. A survey by FreightWaves and Trimble highlights current AI adoption, challenges, and future expectations in the industry.