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

Info · 2026-08-24

ArXiv Proposes SCAE Framework: Revolutionizing Process-Level Evaluation for Coding Agents

The ArXiv team introduces the SCAE (Structural Causal Action Estimator) framework for process-level evaluation of coding agents. SCAE distinguishes between action prediction, task uncertainty, and step attribution by leveraging a structural causal model and a replay-based estimator. Experiments on 499 file-localization episodes from 12 repositories demonstrate that SCAE provides a more accurate measure of causal contribution in agent execution, rather than relying solely on semantic relevance.

Info · 2026-08-24

ArXiv Proposes SPARE Framework: Optimizing Reasoning Pruning in Multimodal Agents

The ArXiv team introduces SPARE, a framework for optimizing reasoning pruning in multimodal agents. Using Kullback-Leibler (KL) divergence, SPARE identifies and removes redundant reasoning text while preserving critical visual evidence, mitigating the impact of 'textual debt' on model performance. Experiments demonstrate that SPARE achieves the highest average accuracy across multiple multimodal tool-use benchmarks while removing 37.89-64.58% of reasoning tokens, showcasing its potential to enha

Info · 2026-08-24

ArXiv Introduces ASP Framework: Optimizing Perception Efficiency for Multimodal Agents under Resource Constraints

The ArXiv team introduces the ASP framework to address the challenge of perception efficiency for multimodal agents under a fixed per-decision token budget. ASP combines a capped structured state, a verbatim episodic index, and query-conditioned budget allocation with iterative access to enhance efficiency. Experiments show that under a 4,096-token decision budget, ASP achieves an episodic retrieval accuracy of 75% to 94%, outperforming traditional query-independent sampling methods.

Info · 2026-08-24

ArXiv Releases Meta-Moderator Framework: Revolutionizing Meta-Cognitive Regulation in Multi-Agent Debates

The ArXiv team introduces Meta-Moderator, a novel framework designed to address the performance bottlenecks in multi-agent debates caused by weak moderation. Meta-Moderator dynamically monitors debate utility, controls the deliberation process, and adjudicates the final answer, thereby making debate regulation an explicit capability rather than an incidental effect of prompting. Trained independently of the debaters via outcome-driven policy optimization, Meta-Moderator outperforms widely used d

Info · 2026-08-24

ArXiv Releases MobilePA-Bench: A Benchmark for Evaluating Mobile Planning Agents on Complex Real-World Tasks

The ArXiv team introduces MobilePA-Bench, a novel benchmark designed to address critical shortcomings in evaluating mobile planning agents. This interactive, stateful, and tool-centric platform spans 13 functional domains and 212 realistic mobile tools, evaluating agents on advanced dimensions such as sub-agent collaboration, memory usage, and skill usage. Experiments show that current frontier LLMs are unreliable in mobile settings, while MobilePA-Bench serves as both a diagnostic tool and a fo

Info · 2026-08-24

ArXiv Releases 'Artificial Empathy' Framework: Revolutionizing AI Agent Recognition and Modeling in Environments

The ArXiv team has introduced a novel framework called 'Artificial Empathy,' which aims to address the challenge of AI systems identifying and modeling other agents in complex environments. This framework leverages a reinforcement learning agent trained on an independent task as prior knowledge to perform agency detection and policy reconstruction based on observed behavioral patterns. This advancement enhances AI's ability to achieve cooperative behavior in real-world scenarios and provides a n

Info · 2026-08-24

ArXiv Releases AutoSaddler: Automatic Harness Optimization for Robust Agent Execution Traces

The ArXiv team introduces AutoSaddler, an automatic optimization framework designed to address the reliability issues of Large Language Model (LLM) agents in long-horizon tasks caused by the compounding of small local failures. AutoSaddler employs offline learning to iteratively update the agent's execution harness using failure signals, resulting in significant performance improvements. Experiments on GAIA2, SWE-Bench Pro, and Terminal-Bench 2.0 demonstrate gains of 9.0, 9.6, and 10.0 percentag

Info · 2026-08-24

ArXiv Releases NIS-Agent Framework: Addressing Inertia Bias in Deep Research Agents

The ArXiv team introduces NIS-Agent, a novel framework designed to address the inertia bias problem in deep research agents. Inertia bias occurs when an agent becomes less objective in evaluating the consequences of its own prior actions, such as queries, plans, or intermediate conclusions. NIS-Agent applies context isolation at two critical decision points—page triage and final-answer validation—to enhance the agent's objectivity and performance. The framework reduces token costs by 33% and dem

Info · 2026-08-24

ArXiv Proposes Reinforcement Learning-Based Macro-Action Topological Navigation for 3D Environment Localization

The ArXiv team has developed a reinforcement learning-based agent navigation method that enables agents to estimate their position and navigate to multiple target objects in large, photorealistic 3D apartment environments using only camera images. By storing ORB features on objects and combining them with an Extended Kalman Filter (EKF) for position estimation, this method replaces traditional SLAM techniques, achieving more efficient navigation. The research demonstrates the agent's ability to

Info · 2026-08-24

Study Reveals: Small Language Model Agents Repeat Failed Calls Due to Flawed Feedback Mechanism

ArXiv researchers have uncovered a behavioral pattern in small language model agents where they tend to repeat failed tool calls instead of learning from errors. This issue stems from the flawed design of the feedback mechanism, where error messages are incorrectly treated as corrective information rather than negative signals to avoid. The study proposes several remedies, such as replacing verbatim calls with runtime-generated failure descriptions and making previously failed strings unreachabl

Info · 2026-08-24

ArXiv Introduces PolyChirp: Multi-Species Birdsong Classification Using TinyML

The ArXiv team has developed PolyChirp, a novel approach that combines biological domain expertise, automated dataset curation, neural architecture optimization, and novel hardware to achieve multi-class bird species detection in the wild. Based on newly designed tiny multiclass models and leveraging recent microcontrollers with neural processing units (NPUs) for hardware acceleration, PolyChirp not only outperforms state-of-the-art single-species binary classification but also robustly classifi

Info · 2026-08-24

ArXiv Releases NetConfArena: Executable Benchmark for Closed-Loop Network Configuration with LLM Agents

ArXiv introduces NetConfArena, an executable benchmark for evaluating the reliability of Large Language Model (LLM) agents in closed-loop network configuration tasks. NetConfArena places agents in emulated multi-device networks, offering a standardized action interface for task execution and evaluating network behavior with hidden task-specific test cases. Built on an LLM-assisted, emulation-grounded pipeline, it converts human-oriented network materials into reusable parameterized task template

Info · 2026-08-24

ArXiv Releases TOWN-VLA Framework: Revolutionizing Smart Intervention in Vision-Language-Action Policies

The ArXiv team introduces TOWN-VLA (Think Only When Needed), a novel framework designed to optimize smart intervention in Vision-Language-Action (VLA) policies. By separating candidate generation from the permission to alter policy inputs, TOWN-VLA prevents performance degradation caused by changes in instruction form. Experimental results demonstrate significant improvements in task success rates across multiple benchmarks. For instance, in the LIBERO-Plus evaluation, the success rate increased

Info · 2026-08-24

ArXiv Releases EvoWiki: An Incremental QA Architecture for Cross-Meeting Knowledge Evolution

ArXiv introduces EvoWiki, an incremental question-answering architecture designed to address the challenges of cross-meeting knowledge evolution. EvoWiki decouples offline incremental construction (Build) from online structured reading (Read), using entity version chains and a fine-grained State-Overwrite Protocol to distinguish current valid states from outdated history while preserving meeting-level provenance anchors. Experiments demonstrate that EvoWiki significantly improves macro-average J

Info · 2026-08-24

ArXiv Releases Apodex 1.1: Revolutionizing Agentic Intelligence for Complex Work

ArXiv has released Apodex 1.1, an agentic framework designed to address the challenges of sustained interaction and state maintenance in complex tasks. The framework enhances an agent's ability to handle files, information sources, and executable code through two complementary dimensions: Environment Scaling and Agentic Coordination Scaling. Experiments demonstrate its superior performance in professional work, finance, scientific research, and more. The lightweight Apodex 1.1 Mini retains stron

Info · 2026-08-24

ArXiv Proposes VAT Determination Agent Decomposition Framework: Exploring Optimal Task Allocation Strategies for LLM Age

ArXiv has released a study on the decomposition of LLM agents, comparing the performance of single-agent and multi-agent architectures in cross-border VAT determination tasks to explore optimal task allocation strategies. The research found that medium-grained agent configurations performed best in terms of accuracy but did not meet the pre-stated performance bar, while the single agent did not Pareto-dominate the orchestrated set. The study also evaluated robustness under different failure inje

Info · 2026-08-24

ArXiv Releases MediSkill-Evo: Revolutionizing Clinical Agent Decision-Making and Knowledge Evolution

The ArXiv team introduces MediSkill-Evo, a clinical agent framework designed to address challenges in evidence-grounded, safe interactions under partial observability. By self-evolving clinical, procedural, symbolic, and visual knowledge in four typed knowledge banks and leveraging a process-constrained preference mechanism, MediSkill-Evo achieves significant improvements in diagnosis accuracy and treatment-intent coverage while substantially reducing critical failures. Experimental results demo

Info · 2026-08-24

ArXiv Introduces SkillAlchemy: Revolutionizing Open-World Agent Skill Creation

The ArXiv team introduces SkillAlchemy, a novel framework for open-world agent skill creation. SkillAlchemy addresses the challenge of generating reliable skills from underspecified briefs by identifying implicit requirements, filtering candidate procedures based on evidence-supported scope, and compiling the admitted content into a grammar-guided skill package. Experiments across 87 SkillsBench v1.1 tasks demonstrate that SkillAlchemy improves the pass rate by 19.9 percentage points over no-ski

Info · 2026-08-24

ArXiv Introduces Reward-Free Continual Learning Framework for Adaptive Control of Space Robots

Space robots operate in extreme environments where hardware degradation can severely compromise traditional control strategies. The ArXiv team introduces a reward-free continual learning framework that leverages latent-state world models. By pre-training a model-based agent across diverse simulations, the world model learns a robust predictor of the reward structure within its latent space. Upon deployment in an environment with severe hardware degradation, the observation encoder and reward pre

Info · 2026-08-24

ArXiv Introduces BasketballBench: A Multimodal Benchmark for Comprehensive Basketball Understanding

ArXiv introduces BasketballBench, a multimodal benchmark designed to address the limitations of existing basketball understanding tasks that evaluate abilities in isolation. The benchmark comprises 7,980 questions across ten tasks in text, image, and video, built from the 2025-2026 NBA season data, including play-by-play, rosters, player profiles, and 2,501 possession-level broadcast clips. Experiments show that current MLLMs struggle with questions requiring the integration of multiple capabili

Info · 2026-08-24

ArXiv Introduces AgentRoom: A CRDT-Backed Multi-Agent Collaborative Coding Protocol

The ArXiv team introduces AgentRoom, a novel real-time collaborative editing protocol designed to address coordination challenges in multi-agent coding. Leveraging Conflict-free Replicated Data Types (CRDTs), AgentRoom enables efficient collaboration among multiple agents on a shared filesystem through mechanisms like file-level claims, status broadcasts, and MCP tools. Experiments demonstrate that AgentRoom outperforms single-agent setups in terms of task abandonment rates and run-to-run stabil

News · 2026-08-24

ArXiv Proposes Hierarchical O-RAN Control Framework for Optimizing UAV-gNB Network Performance

ArXiv researchers introduce a novel hierarchical O-RAN control framework to optimize the performance of UAV-mounted 5G mmWave base stations (gNBs). The framework leverages a Non-Real-Time RIC rApp and a Near-Real-Time RIC xApp to jointly control UAV placement and resource allocation, enhancing the quality of service for enhanced Mobile Broadband (eMBB) and Ultra-Reliable Low-Latency Communication (URLLC). Experimental results demonstrate a 17% improvement in eMBB SLA satisfaction and a 42% incre

News · 2026-08-24

ArXiv Introduces ExTS Tree Search Policy to Optimize Budget-Constrained Agentic Search Efficiency

The ArXiv team introduces ExTS, a novel tree search policy designed to address budget-constrained agentic search challenges. ExTS integrates discriminative reward shaping, a stochastic virtual child, and quality-conditioned branching to significantly enhance search efficiency. In tasks such as prompt optimization, code generation, molecular structure elucidation, and agentic workflow optimization, ExTS demonstrates an average performance gain of 5.5% over task-specific baselines. The study also

News · 2026-08-24

Hugging Face Releases SWE Refactor Bench: A New Benchmark for Evaluating AI Agents in Code Migration

Hugging Face has released SWE Refactor Bench, a benchmark designed to evaluate AI agents' capabilities in code migration tasks. The benchmark consists of 20 whole-repository migration scenarios covering four types of technical debt and employs a three-stage evaluation protocol—Migration Audit, Behavioral Tests, and Agentic Verification—to assess both migration completeness and behavioral correctness. Experimental results show that current AI agents struggle with complex migration tasks, with onl

News · 2026-08-24

Hugging Face Proposes New FSM-Based Approach for Analyzing LLM Agent Behavior

Hugging Face's research team introduces a novel approach using Finite-State Machines (FSM) to analyze the behavior of LLM-based agents. By collapsing unstructured agent traces into compact FSMs, the method achieves significant advancements in cross-task prediction and failure detection. Experiments across twelve public datasets demonstrate the FSMs' efficiency and effectiveness, outperforming existing methods in next-step prediction and failure prediction tasks. This work provides a model-agnost

News · 2026-08-24

Hugging Face Releases AgentRoom: A New Protocol for Concurrent Multi-Agent Coding

Hugging Face has introduced AgentRoom, a real-time collaborative editing protocol for concurrent multi-agent coding. Leveraging Conflict-free Replicated Data Types (CRDTs), AgentRoom enables file-level claim, status, and broadcast functionalities on a shared filesystem, addressing the limitations of existing multi-agent systems in handling parallel tasks. Experiments show that AgentRoom outperforms traditional single-agent approaches by reducing task abandonment rates and run-to-run variations,

News · 2026-08-24

Hugging Face Releases MARS: Revolutionizing Competitive Programming with Multi-Specialist LLM Relay System

Hugging Face has introduced MARS (Multi-Agent Relay of Specialized LLMs), a novel framework designed to address the limitations of existing LLMs in competitive programming. MARS employs retrieval-augmented generation over an algorithm-theory corpus, enabling each agent to specialize in specific topics such as dynamic programming, graph theory, string processing, and geometry. Through collaborative relay, MARS achieves a 62.4% pass rate on the CodeContests test set, a 14.4 percentage point improv

News · 2026-08-24

CyberFactory Releases Aegis: Revolutionizing Cybersecurity with Open-Source AI Framework

CyberFactory has released Aegis, an open-source AI framework designed to enhance cybersecurity capabilities through automation and intelligence. By transforming public vulnerability artifacts into executable task instances and leveraging skill-guided reinforcement learning, Aegis automates the entire process from vulnerability analysis to patch generation. The framework demonstrated strong performance on the CyberGym benchmark, achieving a Pass@1 rate of 52.4% within a one-hour budget, outperfor

News · 2026-08-24

Hugging Face Releases AutoSaddler: An Automatic Harness Optimization Framework for AI Agents

Hugging Face has introduced AutoSaddler, an automatic harness optimization framework designed to enhance the reliability of LLM agents in long-horizon tasks. By formulating harness improvement as an offline learning problem and iteratively updating the harness using failure signals from mini-batches, AutoSaddler significantly boosts agent performance. Experiments on benchmarks like GAIA2, SWE-Bench Pro, and Terminal-Bench 2.0 demonstrate gains of 9.0 to 10.0 percentage points over baseline harne

News · 2026-08-24

Hugging Face Releases MobilePA-Bench: A Benchmark for Mobile Planning Agents on Complex Real-World Tasks

Hugging Face has released MobilePA-Bench, an innovative benchmark platform designed to evaluate the tool-calling and planning capabilities of mobile planning agents in complex real-world tasks. The platform operates within an interactive sandbox environment, simulating real application scenarios across 13 functional domains and 212 realistic mobile tools. It assesses agents along three advanced dimensions: sub-agent collaboration, memory usage, and skill usage. Experiments show that current fron