#多智能体系统
18 related articles
New Breakthrough in Multi-Agent Collaboration: MASCAL Teaches AI to Express Confidence
MASCAL enables each agent in multi-agent systems to output calibrated correctness probabilities, improving downstream decision quality through confidence-aware weighting and demonstrating generalization across collaboration architectures and tasks.
How Do Multi-Agent Medical Systems 'Launder' Bias? New Study Reveals Unfairness Beneath Consensus
New study reveals "disparity laundering" in clinical multi-agent systems: deliberation cannot eliminate bias but hides it. Learn how diverse-perspective ensembles reduce disparities.
Two RL Agents Team Up to 'Chopstick-Catch a Rocket': Mechazilla Captures Super Heavy Booster in Practice
A developer trains two PPO agents in JAX to control a Super Heavy booster and Mechazilla tower arms for cooperative capture, detailing staged training, rollout length, and learning rate tuning.
SwarmRSI: How a Swarm of Research Agents Achieves Recursive Self-Improvement
The SwarmRSI framework uses a shared graph to enable a swarm of research agents to achieve recursive self-improvement, addressing the two pain points of failed-experiment attribution and wasted intermediate results. It sets new SOTA on OpenProblems, NatureBench, and MLEBench Lite, and its generated programs are already used by biologists for immunity research.
The Challenge of Interaction-Mediated Nonstationarity in Multi-Agent Reinforcement Learning
A Reddit thread explores interaction-mediated nonstationarity in multi-agent RL, analyzing its sources, the limits of traditional methods, and adaptation strategies such as meta-learning, hierarchical learning, and graph neural networks.
How Social Scaffolding Shapes Self-Evolving Multi-Agent Systems
The AGORA framework reveals how social scaffolding influences self-evolving multi-agent systems: team composition and dependency structure determine evolution outcomes, and the right scaffolding can exceed individual evolution baselines by 25 percentage points.
When AI Agents Spawn Themselves: The Accountability Problem in Multi-Agent Systems
When AI agents autonomously spawn sub-agents and modify code, who's accountable? This article examines the identity, delegation, and provenance challenges in multi-agent systems, and why orchestration-layer governance may matter as much as agent capabilities.
SpawnRipple: An External Social Environment Built for Autonomous AI Agents
SpawnRipple is an external social environment for autonomous AI agents — providing identity, publishing, discovery, interaction, and API without running any agent models.
How Does an Agent Decide to Trust Another Agent? The Trust Problem in Multi-Agent Collaboration
When AI Agents start autonomously hiring other Agents, how do they decide who to trust? This article explores the missing portable reputation system in multi-agent collaboration and what solutions might look like.
Deep DivesDeep Dive into Pi's Swarm System: A Biology-Inspired Multi-Agent AI Programming Architecture
Deep dive into Pi's swarm system architecture (26K GitHub stars): scout, worker, and soldier ant roles, pheromone communication, adaptive concurrency control, and how multi-agent collaboration revolutionizes AI programming.
Product ReviewsOpenSwarm Hands-On: Open-Source Multi-Agent System Generates Complete Investment Package from a Single Prompt
Deep dive into OpenSwarm, an open-source multi-agent system where 8 specialized agents collaborate to generate complete deliverables—research, charts, slides—from a single prompt.
TutorialsThe Three-Layer Pyramid Model for Agent Development: From Beginner to Industrial-Grade Deployment
Deep dive into the Three-Layer Pyramid Model for Agent development, covering autonomous agents, collaborative multi-agent systems, and universal orchestration agents with a complete learning path from beginner to industrial-grade deployment.
ResearchGoogle Antigravity Uses 93 AI Agents to Build a Complete Operating System from Scratch
Google Antigravity built a complete OS from scratch using 93 AI agents and a single prompt—including kernel, drivers, and all components—for under $1,000.
Deep DivesHow NVIDIA's Multi-Agent System Automates Financial Signal Discovery
Deep analysis of NVIDIA's multi-agent system for quantitative financial signal discovery, covering architecture design, LLM financial reasoning, and automated backtesting iteration.
Deep DivesFactory Missions Architecture Deep Dive: How a Multi-Agent System Runs Continuously for 16 Days
Deep dive into Factory Missions multi-agent architecture: how validation contracts, serial execution, and structured handoffs enable an AI coding system to run 16 days without losing context.
Tech FrontiersDeepMind Partners with EVE Online: Why This 20-Year-Old Game for AI Research?
Google DeepMind partners with CCP Games for AI research in EVE Online. This space sandbox with complex economics and thousands of concurrent players becomes a new testbed for multi-agent systems and reinforcement learning.
Deep DivesARS Academic Framework Explained: A Complete System Where 35 AI Agents Collaborate on Research
Deep dive into the ARS academic framework: how 35 AI Agents collaborate on literature review, paper writing, and quality assessment with a ten-step workflow, multi-layer quality control, and academic integrity safeguards — all for just $4-6.
Deep Divesbest-agent: A Self-Evolving Multi-Agent Framework Built on Claude Code
Deep analysis of best-agent, a self-evolving AI Agent framework built on Claude Code with 94+ skills, 14 specialized agents, and Computer Use capabilities.