Skip to main content
ZICQ

Topics Agent Frameworks

Agent Frameworks

AI Agent frameworks compared: LangGraph, AutoGen, CrewAI, Letta — design philosophy, tool use model, and production fit.

· Aug 15, 2026 #Agent#agent#LangGraph#LangChain#AutoGen#CrewAI#Letta · 30 / 4352 articles

News · 2026-08-27

BTS-AgentBench Released: A Benchmark Framework for Constructing Executable Multi-Turn Agent Tasks

BTS-AgentBench is a novel benchmark framework designed to transform read-only telemetry data from industrial sites into executable multi-turn agent tasks. The framework standardizes BTS metadata and raw histories into a read-only tool store, compiling tasks with tool-derived gold answers, evidence, and typed, bounded operator-facing episodes. Experiments demonstrate BTS-AgentBench's high consistency and reliability in data processing and task construction, offering a new technical pathway for ag

News · 2026-08-27

ArXiv Proposes MoRe Framework: Enabling Multi-Role Reasoning in a Single Agent

ArXiv recently released a new study named Mixture of Roles (MoRe), which aims to address the limitations of single-agent systems in multi-role reasoning. MoRe learns a diversified codebook of steering vectors and uses a query-aware router to dynamically fuse multiple roles into a single steering vector, enabling multi-perspective problem-solving in a single-agent, single-turn inference process. Experiments show that MoRe outperforms traditional single-agent methods in reasoning efficiency and pe

News · 2026-08-27

INTENT-AS-A-TOOL: A Novel Approach to Track Agentic Misalignment

As large language models (LLMs) are increasingly deployed as autonomous agents, safety concerns arise, particularly regarding harmful actions taken by agents under goal conflicts and pressures. INTENT-AS-A-TOOL introduces intent-targeted tools to provide a dedicated channel for the model to express commitment to target behaviors. The probability of calling an intent tool serves as a judge-free, fine-grained signal of the model's tendency to pursue specific actions. The research demonstrates that

News · 2026-08-27

ArXiv Releases ESRP-Bench: A New Benchmark for Advancing Embodied Agent Scene Rearrangement Planning

ArXiv has introduced a novel task called Embodied Scene Rearrangement Planning (ESRP) and released a comprehensive benchmark, ESRP-Bench, to facilitate research in this area. The task requires embodied agents to rearrange furniture in 3D scenes using only egocentric observations and a top-down target layout, without access to the global state, and with mutual object occlusions to reflect real-world constraints. ESRP-Bench, built on OmniGibson, features over 5,400 scene pairs and 8,200 objects, a

News · 2026-08-27

ArXiv Releases FlashVLA Framework: Enhancing VLA Model Inference Speed and Asynchronous Execution Stability

ArXiv has introduced FlashVLA, a novel streaming action decoding framework designed to address the high inference latency and unstable asynchronous execution challenges faced by Vision-Language-Action (VLA) models in robotic manipulation. FlashVLA maintains a streaming action buffer with multiple noise levels and employs chunk-wise causal attention for decoding, enabling the generation of one executable action chunk per inference step. This design significantly improves control frequency and ens

News · 2026-08-27

UrbanGround Released: A Sandbox for Testing AI Agents in Complex Urban Environments

UrbanGround is a novel sandbox platform designed to test AI agents' perception and action capabilities in complex urban environments. Built on Hong Kong's 3D geospatial data, it supports first-person view closed-loop interactions and provides an interactive map for navigation, enabling agents to explore a virtual replica of a real city. Through three research questions, UrbanGround evaluates agents' abilities in scene understanding, navigation, and handling dynamic changes. The study reveals tha

News · 2026-08-27

ArXiv Introduces RTNav: Enhancing Real-Time Zero-Shot Object Navigation

ArXiv introduces RTNav, a novel architecture designed to address the real-time challenges in zero-shot object navigation. By explicitly accounting for inference latency, asynchronous environment stepping, and bounded compute, RTNav significantly improves task performance. In evaluations across HM3D-v1, HM3D-v2, and HM3D-OVON benchmarks, RTNav achieves up to an 11% increase in success rate and up to 5.1 points in Success weighted by Completion Time over prior methods. This research offers a new t

News · 2026-08-27

ArXiv Releases PILOT Framework: Enabling Real-Time Self-Improvement and Task Redirection for Agents

ArXiv has released PILOT, a novel framework designed to enable real-time self-improvement and task redirection for intelligent agents. PILOT employs two coupled mechanisms—live steering and live self-evolution—to allow agents to dynamically adjust their trajectories and optimize performance during task execution. Experimental results demonstrate that PILOT outperforms traditional methods across multiple benchmarks, achieving up to a 9.8 percentage point improvement on Terminal-Bench 2.0, while s

News · 2026-08-27

ArXiv Releases DuMateBench: Benchmark for Evaluating Autonomous Agents in Complex Real-World Workflows

ArXiv has introduced DuMateBench, a novel benchmark designed to evaluate autonomous agents in complex real-world workflows. By reconstructing anonymized and privacy-screened user sessions from a large-scale production agent platform, DuMateBench simulates real-world task complexity, environmental noise, and interaction history. The benchmark comprises 200 tasks across 8 broad scenarios and 17 fine-grained capability categories, using a hybrid evaluation protocol to assess agent performance under

News · 2026-08-27

ArXiv Introduces Safe-CRL: Enhancing Safety and Performance in Goal-Conditioned Reinforcement Learning

ArXiv has introduced Safe Contrastive Reinforcement Learning (Safe-CRL), a novel approach to address the overestimation bias in traditional goal-conditioned reinforcement learning (CRL) within failure-terminated Markov decision processes. By incorporating mass-weighted InfoNCE and a log-survival-mass score, Safe-CRL corrects the bias and enhances safety. Experiments across twelve failure-prone robot navigation and locomotion tasks demonstrate that Safe-CRL significantly improves survival rates a

News · 2026-08-27

ArXiv Introduces proFILL: A Profile-Aware RAG System for University Academic Advising

ArXiv introduces proFILL, a novel method that transforms the rule-based college academic advising chatbot hoBIT into a profile-aware retrieval-augmented generation (RAG) system. proFILL progressively acquires only the necessary user profile attributes for each query, guided by both the query intent and the initially retrieved evidence, to optimize the retrieval process for more accurate answers. Extensive experiments and a human preference study demonstrate that proFILL outperforms diverse RAG b

News · 2026-08-27

ArXiv Releases AgentJudgeBench: First Benchmark for Evaluating LLM Judges in Agentic Tool-Calling

ArXiv introduces AgentJudgeBench, the first benchmark designed to systematically evaluate the reliability of Large Language Models (LLMs) acting as judges in agentic tool-calling systems. The benchmark analyzes LLM performance across varying levels of difficulty and DAG topologies, revealing significant limitations in handling complex, dependency-driven workflows. The study finds that task difficulty is the primary factor impacting judge alignment, with alignment degrading faster in the absence

News · 2026-08-27

ArXiv Releases Framework for Risk and Control in Multi-Agent Systems: Addressing AI Interactions Across Organizational B

ArXiv has released a research report presenting a framework for analyzing risks in multi-agent systems, particularly focusing on AI agent interactions across organizational boundaries. The framework categorizes interactions into three tiers—singular governance, federated governance, and open environments—and examines risk factors, failure modes, and available controls for each. This work underscores the growing complexity of AI agent interactions and emphasizes the need for robust governance and

News · 2026-08-27

ArXiv Introduces SIGMA Framework: Enhancing Robustness of Multi-Agent Reinforcement Learning in Noisy Environments

ArXiv introduces SIGMA, a novel hierarchical collaboration framework for multi-agent reinforcement learning (MARL) that enhances robustness under noisy observations. By organizing agents into adaptive local structures through density-based grouping and performing intra-group consensus aggregation, SIGMA preserves shared task-relevant information while smoothing agent-specific deviations. Inter-group attention further integrates information across groups to maintain global coordination. Experimen

News · 2026-08-27

ArXiv Proposes Five Primitives to Revolutionize Runtime Governance of Autonomous AI Agents

ArXiv has released a research paper proposing five core primitives—Discovery, Identity, Governance, Attestation, and Supply Chain—to address the challenges of governing autonomous AI agents at runtime. These primitives aim to resolve the shortcomings of traditional control models when dealing with ephemeral, model-driven agents whose actions are not pre-programmed. The research describes an implementation that mediates agent actions against policy and records them in a hash-linked signed ledger

News · 2026-08-27

ArXiv Releases Co-Scientist: A Gemini-based Multi-Agent System for Accelerating Scientific Research

ArXiv has released a study on Co-Scientist, a Gemini-based multi-agent system designed to accelerate end-to-end scientific research from hypothesis generation to experimentation and manuscript writing. This specialized system goes beyond in silico simulations, interfacing with real-world lab equipment to design safe precursor routes for MXenes and optimize growth recipes for semiconductors. Additionally, Co-Scientist predicts emergent swarming behaviors in biology and discovers novel inference-t

News · 2026-08-27

ArXiv Releases LiveVVT: Enabling High-Fidelity Real-Time Video Virtual Try-On

ArXiv has introduced LiveVVT, a novel real-time video virtual try-on (VVT) technology that leverages a rolling streaming diffusion framework to maintain local bidirectional spatio-temporal modeling while significantly reducing latency and increasing throughput. By jointly denoising multiple video chunks within a fixed-size window and employing complementary memory mechanisms for long-term consistency, LiveVVT achieves superior generation quality compared to similarly sized models, with 26x lower

News · 2026-08-27

ArXiv Releases Daydreaming: Stealing Hidden Agent Skills via Black-Box Task Interaction

ArXiv introduces Daydreaming, a novel attack method that steals multi-file skills through black-box task interactions without directly requesting the skill or reconstructing its text. By adaptively crafting tasks to distinguish hidden behaviors, Daydreaming recovers 86.8% of the original skill's capability across 7 skills and 4 victim models, outperforming SigLeak by nearly 4x. This research highlights the limitations of current disclosure defenses and underscores the risks of functional reconst

News · 2026-08-27

ArXiv Introduces AI Control Scientist: Automated Control Design Enters the Agent-Driven Era

ArXiv's latest research introduces AI Control Scientist (AICS), a novel agentic system that leverages large language models (LLMs) to automate the generation of optimized controllers from language-based design requirements. This system interprets user requirements, generates candidate controller structures, and refines parameters based on closed-loop performance criteria. Experiments demonstrate that AICS outperforms existing automated baselines in design success rate and optimization efficiency

News · 2026-08-27

ArXiv Introduces Instruct-to-Act Framework: Decoupling Planning and Control for Efficient Agent Behavior

ArXiv introduces the Instruct-to-Act framework, a novel approach that decouples high-level planning from low-latency control to achieve efficient agent behavior. By leveraging pre-trained vision-language models (VLMs) for generating sparse, high-level text instructions and training a world-model controller for fast, autonomous actions, the framework excels in diverse environments. Evaluations across seven embodied environments, including multi-agent scenarios, demonstrate its superior performanc

News · 2026-08-27

ArXiv Releases LiveSim: Revolutionizing User Behavior Simulation in Live-Stream Ecosystems

ArXiv has introduced LiveSim, an innovative framework for simulating user behavior in live-stream ecosystems. Leveraging large language models (LLMs), LiveSim represents users as editable behavioral hypotheses and iteratively refines them through trajectory-grounded interactions to capture the environmental shaping effects on user behavior. Experiments demonstrate LiveSim's effectiveness in improving behavioral fidelity and enabling ecosystem-level analysis of risk evolution and platform interve

News · 2026-08-27

ArXiv Proposes RL-Based Framework for CAV Platoon Joining Maneuvers in Mixed Traffic

ArXiv presents a novel reinforcement learning (RL)-based framework for investigating and optimizing Connected and Automated Vehicle (CAV) platoon joining maneuvers in mixed traffic environments. The research focuses on achieving safe and efficient maneuvers by incorporating risk penalties into the learning process or using an external safety controller to constrain the learned policy. Results show that Proximal Policy Optimization (PPO) outperforms other algorithms like DQN and DDQN, achieving a

News · 2026-08-27

ArXiv Releases DSA Framework: A New Paradigm for LLM-Agent Orchestration in Multi-Market Stock Research

ArXiv introduces DSA (Evidence-Aware Orchestration Framework), an innovative solution for orchestrating Large Language Model (LLM) agents in multi-market stock research. DSA structures the workflow into evidence acquisition, context construction, model-routed analysis, and report generation, incorporating role and strategy skill reasoning to optimize decision-making. The framework supports regional market paths, bundled strategy skills, and both hosted and local model routes, offering a more eff

News · 2026-08-27

ArXiv Introduces ASIL: Revolutionizing Agent-Software Interaction

ArXiv introduces ASIL (Agent-Software Interaction Layer), an innovative interface that replaces the traditional screenshot-and-click approach with structured JSON observations and executable semantic actions. ASIL leverages the deepest feasible access path for each application and has been instantiated across 15 applications, demonstrating superior performance in 300 single-application and 80 multi-application tasks. Compared to screenshot-and-click control, ASIL achieves higher task success rat

News · 2026-08-27

ArXiv Releases Multi-Modal AI Framework for Optimizing Real-Time Queue Management in Intelligent Border Control

ArXiv introduces a novel multi-modal AI framework tailored for intelligent border control systems. By integrating heterogeneous data sources and employing LSTM networks for real-time queue prediction, scheduling optimization, and resource management, the framework significantly enhances border control efficiency. Experimental results demonstrate a 35% reduction in queue prediction error, a 30% decrease in average waiting time, and nearly a 20% increase in average throughput compared to tradition

News · 2026-08-27

ArXiv Proposes Soft Active EMG Interface for Machine Learning-Enabled Silent Speech Recognition

ArXiv has introduced a novel soft active electromyography (EMG) interface for silent speech recognition (SSR), leveraging machine learning for efficient word-level speech recognition. The device, worn on the hand, uses a fingertip electrode to capture EMG signals when needed. It integrates liquid metal interconnects, transparent flexible printed circuit electrodes, and elastomer encapsulation to ensure signal stability and mechanical reliability. Experiments demonstrate an average accuracy of 97

News · 2026-08-27

ArXiv Research: Breakthrough in Goal-Priority Regulation via Emotional Preferences

ArXiv has published a study on decision-making for agents, introducing a computational framework for regulating goal priorities based on emergent emotional preferences. The research formalizes the dynamic prioritization between high-level and low-level objectives using reinforcement learning and proposes a framework consisting of a multi-objective reinforcement learning inner controller and an outer preference generator. Experiments in self-constructed multi-objective environments demonstrate th

News · 2026-08-27

ArXiv Introduces Robotic Platform for Active Sensing-Based Plant Stress Characterization

ArXiv has introduced a novel robotic platform designed for targeted fluorescence measurements on plant leaves. This system integrates 3D plant reconstruction, geometric analysis, and motion planning to localize suitable measurement points and generate collision-free trajectories for a robotic manipulator. By enabling automated, repeatable, and spatially resolved physiological measurements, the platform advances beyond passive imaging and offers a robotics-driven approach to high-resolution plant

News · 2026-08-27

ArXiv Releases LoopHarness: Addressing Cross-Iteration Safety State for Autonomous LLM Agents

ArXiv has released a study on the safety of large language model (LLM) agents, introducing a novel approach called LoopHarness to address the issue of cross-iteration safety states. Traditional safety mechanisms reset their state at the start of each new trajectory, making them vulnerable to attacks that span multiple iterations. LoopHarness restores a persistent, non-decaying safety state at the loop level, significantly enhancing the agent's defense against complex attacks. The research also p

News · 2026-08-27

ArXiv Proposes SARA Framework: Separating Action Induction from Execution Authorization in Tool-Augmented LLM Agents

ArXiv has released a study on tool-augmented Large Language Model (LLM) agents, proposing a novel framework named SARA (Separating Action Induction from Execution Authorization). SARA distinguishes action induction from execution authorization by using context-isolated action probes and execution authorization mechanisms, reducing the risk of tool outputs causing unintended real-world side effects. The framework significantly lowers the Action Specification Rate (ASR) across multiple benchmarks