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 / 4351 articles

News · 2026-10-14

arXiv Introduces Agentic Data Engineering Framework: Revolutionizing LLM Data Engineering

arXiv has introduced the Agentic Data Engineering (ADE) framework, a novel approach to LLM data engineering that leverages agents to iteratively improve strategies based on experimental feedback. The framework demonstrates significant performance gains across tasks such as math and code data selection, math reward design, and math curriculum learning, achieving up to 13.34% absolute improvements over generic baselines under fixed training conditions. The ADE framework highlights the potential of

News · 2026-10-14

arXiv Releases Hybrid Reinforcement Learning Trading Framework: Enhanced Risk-Adjusted Returns with Financial Sentiment

arXiv introduces a novel hybrid reinforcement learning trading framework that integrates intraday market data with Transformer-based financial sentiment signals. By extracting sentiment dynamics and semantic event clusters from financial news and incorporating them into a PPO-based trading agent, experiments on NSE intraday data demonstrate significant improvements in risk-adjusted returns compared to price-only baselines.

News · 2026-10-14

arXiv Introduces IDM: Revolutionizing Predictive Mechanisms for Self-Evolving Agents with Task-Level Improvements

arXiv introduces a novel approach to self-evolving agents with the Invertible Improvement Dynamics Model (IDM). IDM predicts task-level outcomes of proposed modifications before costly evaluations, guiding optional revisions and accumulating predictive knowledge across generations. Evaluated across multilingual coding, paper review, robotics reward design, and Olympiad-level mathematics grading, IDM outperforms existing methods by up to 46% in average performance. The model's predictive accuracy

News · 2026-10-14

Google DeepMind and Singapore GovTech Launch Lightweight AI Scam Detection Model

Google DeepMind and Singapore's GovTech have collaborated to develop a lightweight AI scam detection system based on the Gemma model. This on-device solution intercepts conversational scams in private chats in real-time without sending data to the cloud, ensuring zero data egress and enhancing user privacy and security. This collaboration represents a significant advancement in AI-driven security solutions.

News · 2026-10-14

arXiv Releases CIMEMORIES DIALOGUE: Evaluating Privacy Leakage in Agentic Dialogue of Large Language Models

arXiv has released a study on privacy leakage in large language model (LLM) agents, introducing a benchmark called CIMEMORIES DIALOGUE. This benchmark consists of 4,410 two-agent dialogue instances grounded in Contextual Integrity theory, designed to systematically evaluate the privacy protection capabilities of 7 open-weight models from the Qwen3 and Llama 3.1 families. The research reveals a fundamental coupling between task capability and privacy leakage: more capable models are systematicall

News · 2026-10-14

arXiv Proposes New Hybrid Architecture: Enhancing Edge Agent Task Success with Cloud-Written Programs and Local Models

arXiv introduces a novel intermediate architecture that combines cloud-written programs with local model inference to enhance the performance of edge agents. This approach delegates bounded user data-related questions to a small local model through function calls, maintaining data locality while improving task success rates by up to 14.9 percentage points in the AppWorld benchmark. The study demonstrates a significant reduction in the performance gap compared to cloud-based ReAct loops, offering

News · 2026-10-14

arXiv Releases CTI-REALM Benchmark: Open-Weight Models Achieve Breakthrough in Cost-Effectiveness for Security Agents

arXiv has released a study on the CTI-REALM benchmark, evaluating 26 language models (19 open-weight and 7 closed models) in threat intelligence processing tasks. The research reveals that open-weight models can match the strongest closed models' performance at less than a tenth of their cost, provided they are given enough interaction turns. The results also indicate that model cost is not directly correlated with capability, with task completion efficiency being a key differentiator. Additiona

News · 2026-10-14

ClawLoop Framework Released: A Novel Training Loop for Self-Improving Agents

ClawLoop is a novel framework for self-improving agents that introduces a self-sustaining training loop without relying on external teacher models. The framework consists of three phases: task synthesis, agent training (via SFT and/or RL), and diagnosis & feedback. By iteratively refining tasks based on agent performance analysis, ClawLoop consistently enhances agent capabilities across benchmarks like SkillsBench, Claw-Eval-Live, and ClawGym-Bench. The framework's ability to transfer improvemen

News · 2026-10-14

arXiv Releases Analysis of Failure Modes in Agentic Data Science Systems

arXiv has released a research paper analyzing the failure modes in Agentic Data Science (Agentic DS) systems. The study examines 220 data science tasks and identifies prevalent methodological errors such as data leakage and improper validation in existing agentic frameworks and models. A reusable test suite is developed to systematically detect six specific failure modes, and qualitative assessments are conducted against expert solutions from Kaggle competitions. The findings reveal that while a

News · 2026-10-14

NVIDIA AI Agents Power Smarter Urban Operations Across Europe

NVIDIA has announced that its AI agents and digital twin technologies are being deployed across European cities to optimize urban operations and public services. These AI solutions, powered by NVIDIA accelerated computing, NVIDIA Cosmos, NVIDIA Nemotron, and NVIDIA Omniverse, cover areas such as traffic management, energy optimization, facility maintenance, and public safety. Several software partners and system integrators are involved in developing and deploying these projects, highlighting NV

News · 2026-10-14

PsyBer-Agent Released: Psychology-Driven Recommender System Simulator Launched

PsyBer-Agent is a psychology-driven recommender system simulator designed to address the limitations of current LLM-based agents in capturing user dynamics such as fatigue, affect, and shifting preferences. Its core component, the Psy-Engine, models user behavior through three evolving latent states and aligns simulated trajectories with real interaction logs using unbalanced Gromov-Wasserstein optimal transport for structured supervision without the computational burden of fine-tuning the backb

Info · 2026-10-14

Blumify Launches AI Leaders Tracking Tool: Capturing Shifts in Perspectives in Real-Time

Blumify has launched the 'AI Leaders Tracking Tool,' a novel solution designed to analyze video content and track the perspectives of AI leaders and their shifts over time. The tool automatically identifies and categorizes key viewpoints within videos, generating visual trend analysis reports that enable users to quickly grasp the dynamics of the AI landscape and the evolving stances of prominent figures. This innovation offers a new form of data support for AI research, policy-making, and marke

News · 2026-10-14

Google DeepMind and Singapore Government Launch On-Device AI Scam Detection Tool

Google DeepMind and GovTech Singapore have partnered to launch an on-device AI scam detection tool based on the lightweight Gemma model. This tool intercepts conversational scams in private chats in real-time without sending data to the cloud, ensuring zero data egress and enhanced user privacy. The initiative aims to protect citizens from the growing threat of online scams while maintaining data security.

News · 2026-10-13

Hugging Face Explores New Paradigm for Image Generation with Coding Agents

Hugging Face's research team proposes a novel approach to image generation using coding agents that construct images through executable programs, rather than relying on implicit parameter learning in traditional generative models. The study reveals that these agents excel at preserving structure and satisfying precise visual constraints but are limited in perceptual richness. Additionally, code is treated as an external visual language of thought, retaining substantial information for visual que

News · 2026-10-13

Hugging Face Releases Socratic Skill Learning: Revolutionizing Agent Skill Transfer and Learning

Hugging Face introduces Socratic Skill Learning (SSL), a novel approach to enhance the skill transfer capabilities of Large Language Model (LLM) agents. SSL interleaves expert guidance with student practice, enabling agents to question, test, and adapt advice, leading to more efficient skill reuse. The method demonstrates superior performance on the SkillTransferBench benchmark, significantly improving task utility. Additionally, Issue-Guided SSL optimizes the review process by tracking unresolv

News · 2026-10-13

Hugging Face Releases LLM2R: Revolutionizing 3D Part Assembly with Large Language Models

Hugging Face has released LLM2R, a novel approach to 3D part assembly that leverages a pretrained decoder-only large language model (LLM) to predict rigid SE(3) transformations for each part. By compressing part geometries into geometric tokens and processing them with a causal LLM adapted via Low-Rank Adaptation (LoRA), LLM2R achieves state-of-the-art performance in part accuracy, rotation error, and translation error across multiple benchmarks. Its unique architecture, which includes learnable

News · 2026-10-13

Lambda Humanoid Releases HumanVerse and λ_0: Advancing Humanoid Whole-Body Loco-Manipulation

Lambda Humanoid introduces HumanVerse, a 500-hour dataset of open-world human loco-manipulation, and λ_0, a humanoid foundation model for whole-body manipulation. HumanVerse pairs egocentric video with synchronized body and hand motion, addressing the data scarcity in humanoid robotics. λ_0 leverages aligned human demonstrations to improve task performance, achieving a 90.0% success rate on the SIMPLE benchmark and 62.5% across real-world tasks. This work paves the way for scalable human data co

News · 2026-10-13

Hugging Face Releases CyberArena: A Dynamic Adversarial Benchmark for AI Agents in Cybersecurity

Hugging Face has launched CyberArena, a dynamic adversarial benchmark designed to evaluate AI agents in cybersecurity scenarios. Unlike traditional static benchmarks, CyberArena simulates real-world cyber operations through multi-round attack-and-defense combat, testing agents' abilities to simultaneously attack and defend in continuously changing environments. The platform employs an automated environment generation pipeline and a dynamic scoring model to provide a more realistic and nuanced ev

News · 2026-10-13

Hugging Face Proposes Capability-Oriented Attack-Surface Profiling for Agent Skills

Hugging Face has introduced a two-phase attack-surface profiling methodology for agent skills, combining static analysis, dependency provenance checking, instrumented sandbox execution, and model-based trace assessment. The study reveals significant security vulnerabilities: half of the skills exhibit shell-execution capabilities, 35% can read the filesystem, 31% can write to it, 31% can perform network egress, and 13% can access credentials. The existing permission model is largely ineffective,

News · 2026-10-13

Hugging Face Releases Sauron: Revolutionizing AI Agent Verification for Enhanced Task Reliability

Hugging Face has released Sauron, an evidence-first AI agent verifier designed to address the challenges of evaluating long-horizon, semi-verifiable agentic tasks. Sauron independently validates claims against environment states and execution traces, mitigating errors and biases inherent in rubric-based evaluations. Its key mechanisms include bounded queries, joint investigation of related criteria, and auditable receipts for supporting evidence. In multiple benchmarks, Sauron outperformed the s

News · 2026-10-13

Hugging Face Releases EC2-Bench: Benchmark for Evaluating Edge-Cloud Collaboration in Long-Horizon Agent Execution

Hugging Face has released EC2-Bench, a benchmark for evaluating edge-cloud collaboration in long-horizon agent execution. The platform provides executable tasks and a unified evaluation framework covering task performance, resource usage, and privacy exposure, enabling researchers and developers to gain a deeper understanding of edge-cloud collaboration in complex scenarios. Experiments show that collaborative execution achieves a favorable balance among task performance, resource usage, and pri

News · 2026-10-13

Hugging Face Releases BELIEFWEAVER Framework: Enhancing Failure Recovery for Long-Horizon Agents

Hugging Face introduces BELIEFWEAVER, a framework designed to address the challenge of failure recovery in long-horizon agents, where causes of failure may lie several steps back and are beyond the scope of local reflection or backtracking. The framework leverages persistent evidence across attempts and prospective reasoning for repair directions, maintaining a relational belief state and utilizing an implicit world model (IWM) to evaluate repair hypotheses, predict progress, and guide explorati

News · 2026-10-13

Hugging Face Launches 'Into the Unknown': A New Benchmark for AI Agents in Scientific Research

Hugging Face has introduced 'Into the Unknown,' a novel benchmark designed to assess AI agents' capabilities in conducting open scientific research. The benchmark provides agents with scientific papers and research directions, tasking them to write follow-up papers. Collaborating with scientists from U.S. Department of Energy national laboratories, the benchmark establishes evaluation criteria. Although none of the 135 submissions fully met the conditions, and 114 failed the integrity check, thi

News · 2026-10-13

CarryOn Framework Released: Revolutionizing Long-Horizon Agent Context Construction

CarryOn is a novel object-oriented context runtime framework designed to optimize long-horizon agent task state management. By reusing task objects instead of reprocessing interaction histories, CarryOn constructs a context view that maintains task state consistency while significantly reducing token usage. Experimental results on OSWorld 2.0 and BrowseComp-Plus benchmarks show performance improvements of 5.8% and 7.0%, respectively, and a reduction in token usage by 51.9% and 22.9%. This innova

News · 2026-10-13

Hugging Face Releases Cowork-Gym: Revolutionizing Interactive Work Environment Simulation and Training for Agents

Hugging Face introduces Cowork-Gym, a framework for synthesizing and training agents in interactive work environments. The framework employs a build-execute-revise loop, leveraging execution evidence and feedback to validate and repair generated environments, while incorporating a stateful user simulator to capture realistic work dynamics. Experiments demonstrate that Qwen models post-trained with Cowork-Gym outperform their base checkpoints across six agentic benchmarks, particularly excelling

News · 2026-10-13

Hugging Face Releases Harness RSI: Revolutionizing Agent Self-Evolution and Software Architecture Updates

Hugging Face has released the Harness RSI framework, which revolutionizes the self-evolution of agent software architecture by keeping model weights fixed while allowing dynamic revisions to prompts, memory, workflows, and executable components. The framework introduces a feedback-driven revision persistence mechanism, distinguishing between changes in interpreted states and revisions to tools, code, repository control logic, and plugin composition. It employs evolutionary optimization for searc

News · 2026-10-13

Hugging Face Releases EdgeWise: Intelligent Edge-Cloud Collaboration for Long-Horizon Agent Tasks

Hugging Face introduces EdgeWise, an intelligent edge-cloud collaboration framework designed to optimize the execution efficiency of complex, long-horizon agent tasks. By deploying a lightweight SLM on the local device as the primary solver and selectively delegating critical steps to a cloud LLM when stronger model capacity is needed, EdgeWise maintains performance while significantly reducing inference costs. Experimental results demonstrate its effectiveness, with inference costs reduced by u

News · 2026-10-13

Hugging Face Releases CLASHarness: A New Benchmark for Multi-Agent Conflict Generation and Evaluation

Hugging Face introduces CLASHarness, a framework for generating verifiable multi-agent conflicts in computer use scenarios. It addresses the complexities of agent collaboration by combining hierarchical scenario planning with an eight-step verified micro-pipeline to ensure task solvability. The framework includes a benchmark of 600 multi-round tasks. Experiments reveal critical bottlenecks in existing models, such as blind reliance on misleading information, difficulty in fixing system states, a

News · 2026-10-13

Hugging Face Proposes Game-Theoretic Strategic LLM Selection Framework

Hugging Face introduces a game-theoretic strategic decision-making framework for LLMs, redefining strategic selection as a meta-game layered on top of the base game. The framework employs a strategic selector to choose an LLM from a portfolio to navigate a latent, online-learned payoff matrix. Evaluations across five games and ten LLMs reveal that the evaluation signal is game-specific, requiring a higher number of seed runs for reliability. While a lightweight multi-armed bandit algorithm perfo

News · 2026-10-13

Together Computer Releases Hybrid Reasoning Agents: Bridging Latent Reasoning and Text-Based Interaction in AI Agents

Together Computer has unveiled Hybrid Reasoning Agents (HybridAgents), a novel architecture that bridges the gap between latent reasoning and text-based interaction in AI agents. By leveraging a lightweight HybridLink, HybridAgents dynamically navigate between latent and textual modes, while a continuous working memory preserves hybrid context across turns. This design maintains tool-integrated capabilities and ensures transparency of latent thoughts. Evaluated across 11 benchmarks in mathematic