Skip to content
← Daily Digest

Multi-Agent

多 Agent 協定、角色分工、MCP、A2A、通訊架構

Paper Timeline (31)

2026-05-25
SARC: 把法規義務編譯成執行時約束 Must Read

與其把「請遵守法規」寫在 prompt 裡然後希望 AI 記得,SARC 把規範編譯成四個硬性卡關站嵌進代理人的執行迴圈,不依賴 AI 的「記憶力」。

arxiv: 2605.07728
2026-05-25
約束漂移:LLM 多代理系統被忽視的安全失效模式 Must Read

多代理系統中的安全規範不會自動持續有效——它們會在記憶存取、任務委派、工具呼叫等七個環節中悄悄弱化,這篇命名為「約束漂移」並提出分類框架。

arxiv: 2605.10481
2026-05-25
隱形協調者讓多代理系統喪失保護性行為 Skim

用「有沒有可見的領導者」做 AI 實驗:當多代理系統的協調者是隱形的(工作代理不知道有人在指揮),整個系統的保護性行為顯著下降——對流行的 orchestrator 架構是直接的設計警示。

arxiv: 2605.13851
2026-05-26
Predictive Maps of Multi-Agent Reasoning: A Successor-Representation Spectrum for LLM Communication Topologies Skim

借用神經科學的「後繼表示」矩陣,用三個數學量在部署前預測 multi-agent 系統選 Chain、Star、Mesh 哪種拓撲時會出現觀點漂移、假共識或不穩定三種失敗模式。

arxiv: 2605.11453
2026-05-28
AgentFugue: Agent Scaling for Long-Horizon Tasks through Collective Reasoning Must Read

讓多個 agent 同時跑同一個任務,並透過一個「共享推理筆記本」互相學習彼此的發現,比單一強 agent 更能搞定複雜長任務。

arxiv: 2605.24486
2026-06-08
Beyond tokens: a unified framework for latent communication in LLM-based multi-agent systems Skim

Multi-agent 系統裡 agents 之間傳文字很貴又有資訊損失,這篇 survey 整理「改傳 embedding / hidden state / KV-cache」的研究現狀,提出分類框架,幫你評估這條路能否用在你的 agent 系統。

arxiv: 2606.05711
2026-06-10
Autonomous Incident Resolution at Hyperscale: An Agentic AI Architecture for Network Operations Must Read

一套分層多 Agent 系統在真實雲端環境自動偵測、診斷並修復網路故障,生產部署後達 90%+ 自主解決率,不需人工介入。

arxiv: 2606.09122
2026-06-14
Agentic Environment Engineering for Large Language Models: A Survey Must Read

把「怎麼設計 AI Agent 執行環境」系統化:8 個屬性 × 8 個應用領域分類框架,加上 symbolic vs. neural 兩種自動合成範式,是 agent environment 設計的入門地圖。

arxiv: 2606.12191
2026-06-16
Agents All the Way Down: A Methodology for Building Custom AI Agents from Substrate to Production Must Read

不想每次框架升級就重寫 agent?這篇給你從 LLM API 底層到上線維護的完整方法論:兩個前提條件(substrate + building blocks)加上三個持續實踐,比「先 pip install langchain」更底層也更持久。

arxiv: 2606.11869
2026-06-23
Beyond Global Replanning: Hierarchical Recovery for Cross-Device Agent Systems Must Read

跨多台設備的 agent 執行失敗時,別急著全部重來——先讓設備端自己試著換個方式(API → CLI → GUI),不行再往上呈報給總指揮。

arxiv: 2606.20487
2026-06-23
Hierarchical Control in Multi-Agent Games: LLM-based Planning and RL Execution Skim

用 LLM 當多 agent 團隊的指揮官決定策略方向,再讓 RL 訓練的底層 agent 負責真正的動作執行——兩者各司其職,比純 RL 或純手工規則都更彈性。

arxiv: 2606.20014
2026-06-24
PlanBench-XL: Evaluating Long-Horizon Planning of LLM Tool-Use Agents in Large-Scale Tool Ecosystems Must Read

測試 1665 個工具下的長程規劃:最強 LLM 在無干擾時成功率只有 52%,一旦部分工具被封鎖更跌至 11%;這是第一個專門測「大量工具 + 工具失效」組合下 Agent 有多脆弱的 benchmark。

arxiv: 2606.22388
2026-06-24
A Technical Taxonomy of LLM Agent Communication Protocols Must Read

MCP、A2A、ACP、ANP 這些 Agent 通訊協定到底差在哪?TU Munich 分析 9 個主流開源協定,給出第一份技術分類法,讓選協定從「哪個比較紅」變成「哪個技術特性符合我的需求」。

arxiv: 2606.19135
2026-06-24
Arbor: Tree Search as a Cognition Layer for Autonomous Agents Skim

AMD 提出 Arbor:把「樹狀搜尋」當作多 Agent 系統的共享工作記憶,讓 Orchestrator、Specialist、Critic 三類 Agent 圍繞同一棵搜尋樹協作,失敗的嘗試也被保留為後續探索的診斷訊號。

arxiv: 2606.12563
2026-06-25
Sakana Fugu Technical Report Must Read

Sakana AI 訓練了一個「學會指揮其他模型」的 0.6B 小型協調器,把 Claude/GPT 等現有 frontier 模型的能力組合出超越任何單模型的表現,在 SWE-Bench Pro 達到 73.7%(公開最強)。

arxiv: 2606.21228
2026-06-27
Composing Verifiable Conceptual Models via Building Blocks: Towards Design-Time Verification of Agentic AI Workflows Must Read

Agent workflow 設計好、上線後才發現 Agent 卡死或邏輯繞圈,這很常見——這篇提出在「設計階段」用 12 條結構規則自動驗證 workflow 設計有沒有問題,概念類似 IDE 的 type check。

arxiv: 2606.21565
2026-06-27
Beyond Static Leaderboards: Predictive Validity for the Evaluation of LLM Agents Must Read

Agent benchmark 的「總分排行榜」換個情境就失準——這篇用 14 組大規模平行實驗證明排名不穩定,並提出以「預測有效性」(in-sample 與 out-of-sample 排名相關性)取代總分均值作為選型指標。

arxiv: 2606.19704
2026-07-01
The Shift to Agentic AI: Evidence from Codex Must Read

OpenAI 用 Codex 真實使用數據,首次大規模量化「agentic AI 如何改變工作」:2026 年上半年用戶增 5 倍、任務規模膨脹 10 倍,且最重要的是,改變已從工程師擴散到法務、研究等非技術職能

arxiv: 2606.26959
2026-07-02
DeLM: Decentralized Multi-Agent Systems with Shared Context Must Read

把 Multi-Agent 系統裡「一個主管分派所有任務」的架構,換成「所有 agent 共讀一個驗證過的進度表、自己搶任務」——SWE-bench 上省 50% 成本,還拿到更高分。

arxiv: 2606.10662
2026-07-07
UA-ChatDev: Uncertainty-Aware Multi-Agent Collaboration for Reliable Software Development Skim

讓 multi-agent 軟體開發框架學會「自我懷疑」:用 token 的信心分數決定哪個 agent 的輸出需要驗證再轉交,防止早期錯誤在 pipeline 中滾雪球。

arxiv: 2607.02186
2026-07-15
Multi-Agent AI Control: Distributed Attacks Hamper Per-Instance Monitors Must Read

多個 AI agent 協同執行惡意任務時,「每個 agent 各自監控」的防禦策略會因為碎片化效應大打折扣;加入明確攻擊規劃者(planner agent),攻擊成功率最高提升 7 倍。

arxiv: 2607.07368
2026-07-16
PalmClaw: A Native On-Device Agent Framework for Mobile Phones Must Read

現有手機 agent 都靠「模擬點擊螢幕」操作手機,PalmClaw 改成直接呼叫裝置 API:任務成功率提升 11.5%、完成時間縮短 94.9%,而且每一步的執行邊界清楚得多。

arxiv: 2607.13027
2026-07-17
Designing Agent-Ready Websites for AI Web Agents Must Read

用四個設計維度重新打造電商網站,讓 AI 瀏覽器 agent 完成購物任務的成功率從 49% 提升到 89%。

arxiv: 2607.12056
2026-07-20
Branching Policy Optimization: Sandbox-Native Language Agent Reinforcement Learning Skim

現有 RL 訓練算法(PPO / GRPO)對 agent 來說很浪費:每次都從頭跑 N 條路徑比較。BPO 只在「真正需要比較」的高熵決策點快照並分叉,用兄弟軌跡(sibling trajectory)的 return 差異計算 advantage,讓訓練更有效率。

arxiv: 2607.14171
2026-07-20
Orchestrating Power Grid Studies with Multi-Agent AI and MCP Servers Skip

Position paper:以電網研究為案例,展示如何用 MCP(Model Context Protocol)把 LLM agents 接上專業領域仿真工具,並在嚴格需要 human-in-the-loop 的工業流程中落地 multi-agent AI。

arxiv: 2607.14158
2026-07-21
SearchOS-V1: Towards Robust Open-Domain Information-Seeking Agent Collaboration Must Read

搜尋型 agent 老是卡迴圈?SearchOS 把搜尋進度變成「顯性的共享狀態」,加上失敗記憶機制,讓 multi-agent 協同搜尋比 baseline 快 24%、F1 提升 4.3 分。

arxiv: 2607.15257
2026-07-23
MAG: A Web-Agent Benchmark and Harness for Multimodal Action and Guide Generation Must Read

首個把「完成網頁任務」和「生成操作教學」合為一個 benchmark 的多模態評測,完全基於截圖;目前最強模型任務完成率低於 40%,GRPO 訓練讓 9B 小模型完成率從 6.9% 近乎翻倍到 13.2%。

arxiv: 2607.10079
2026-07-29
A Comparative Study of MCP and A2A for Inter-Agent Coordination in LLM-Based Systems Must Read

MCP 管「模型接工具」,A2A 管「Agent 跟 Agent 講話」,兩者不是競爭關係而是分層設計,就像 USB 跟 Wi-Fi 各管各的事,都需要。

arxiv: 2607.23884
2026-07-30
TRACE-ROUTER: Task-Consistent and Adaptive Online Routing for Agentic AI Must Read

多步驟 Agent 任務不能每次呼叫 LLM 都重新選模型;TRACE-ROUTER 改成任務入場時選一次模型並釘死到任務結束,用最終結果回饋訓練選模型的策略,比傳統路由方式在相同延遲下多出 7–8 個準確率百分點。

arxiv: 2607.22465
2026-08-02
ProACT: Towards Breakdown-Aware Proactive Agent in Multi-User Collaboration Must Read

教 Agent 在多人對話中辨識「有人卡住了」,然後選對時機插一句有用的話,而不是一直打擾或一直沉默。

arxiv: 2607.03730
2026-08-03
Toward an Organizational Science of Multi-Agent LLM Systems Must Read

把「哪些 agent 上場」、「怎麼溝通」、「最終怎麼整合答案」三件事分開設計,讓你可以各自換掉,不用每次重頭架系統。

arxiv: 2607.25446

Related Digests (23)

AI Agent GitHub Digest — 2026-08-20

Volcengine (ByteDance) open-sources OpenViking, replacing black-box vector search with a viking:// virtual filesystem for agent memory — benchmarks show 80%+ accuracy while saving 34-91% tokens. munder-difflin wraps multiple coding CLIs into a desktop office with shared memory; ai-memory solves cross-CLI amnesia with a Rust MCP server; mukul975's cybersecurity skill pack rockets to ~28K stars in a day. pydantic-ai v2.32.0 adds OpenRouter/xAI attachment search and instrumentation improvements.

AI Agent Arxiv Digest — 2026-08-18

ActBench red-teams cowork agents via execution traces, finding ASR of 73.7%–94.4% even when swapping harnesses; Agent Behavioral Contracts II shows co-failure rates hit 90% for same-model two-stage pipelines, breaking the conditional independence assumption; Graph-Based RL Drift Diagnosis uses a small-model recovery graph to detect drift and auto-rollback without retraining the primary agent

AI Agent GitHub Digest — 2026-08-16

Vercel ships eve, a filesystem-first TypeScript agent framework tightly coupled with its AI Gateway/Sandboxes; Prime Intellect's Prime Agent treats the entire conversation context as program variables with a self-modifying Continual Harness; aden-hive's Hive replaces pre-compiled execution graphs with 'clone the Queen'; HKUDS's nanobot hits 47k stars in six months with its v0.3.0 Agency Release. No major version bumps on the watchlist today.

AI Agent Arxiv Digest — 2026-08-13

EvoGraph-Mem uses a failure-aware editable graph to let agent memory self-correct, preventing stale insights from poisoning decisions; MAP-Graph turns provenance tracking from post-hoc audit into real-time access control, achieving 95% success across 2,700 synthetic tasks; MaSRead shows multi-agent KV cache sharing is possible but requires content-addressed reading instead of positional addressing

AI Agent Arxiv Digest — 2026-08-11

Muscle Memory proposes 'compiled memory' over retrieval-based memory, winning 88.9% of personalization matchups across 90 scenarios; MoRSE uses role-subtask conditioned LoRA experts to significantly outperform prompt-only role differentiation in code generation; ASCon builds a unified failure attribution model, improving by 5.83%, 10.63%, and 14.73% across three attribution targets

AI Agent Arxiv Digest — 2026-08-05

ToolLIFT abstracts tool trajectories into function-level workflow graphs, lifting OOD accuracy by 4+ points on average; HyperAgent builds tool-schema hypergraphs with deficit-oriented expansion, beating ReAct by 14.3 points on AppWorld with lower token cost; a multilingual multi-agent planning diagnosis finds that planning grounding failures rise with decreasing language resources, and the TART fix improves scores by 5.6 points on average

AI Agent Arxiv Digest — 2026-08-03

Three papers tackling multi-agent platform challenges from three angles: organizational design, security isolation, and user-level authorization. IMACS decomposes multi-agent systems into three independently swappable layers (organization, coordination, collaboration algorithm), letting framework designers mix and match agent roles and strategies like building blocks. APPA uses context branching to break the usability bottleneck of IFC (Information Flow Control), cutting prompt injection exfiltration rates from 31–50% down to 0–7% across 4 models. A UW survey of 21 agent authorization proposals finds that nearly all systems offer only developer-defined global policies — user-level personalized authorization is virtually absent. Together, the three papers outline the gaps agent platforms must close on the road from prototype to production.

AI Agent Arxiv Digest — 2026-08-02

Three papers tackle 'what goes wrong when agents hit production' from different angles: ProACT addresses when an agent should speak up in multi-user collaboration (an Agent UX design problem); the second uses real GitHub data to reveal that coding agents clash with their own PRs (a platform ops pain point); the third surveys five vulnerability classes of cyber-capable agents, using July 2026 HuggingFace/OpenAI incidents as case studies. Together, they form a crash course in post-deployment agent headaches.

AI Agent Arxiv Digest — 2026-07-29

Three papers today converge on infrastructure reliability for production multi-agent systems: the first compares how MCP and A2A divide responsibilities (complementary, not competing); the second benchmarks capability degradation across 12 top models after tool version updates, finding 13-14% drops even in frontier models; the third reveals that chaining safe models into a pipeline does not yield a safe system — defenses actually rely on cloud-provider server-side filters. Together they answer three questions every platform engineer faces: how to connect tools, whether tool upgrades break things, and whether chained agents stay secure.

AI Agent Arxiv Digest — 2026-07-21

Three papers, one question: what makes an agent system actually work? SearchOS-V1 offers an architectural answer — externalize search progress as structured state and record failed paths so multi-agent collaborative search becomes reliable. AutoSynthesis shows that highly structured academic tasks (systematic meta-analysis) can be fully automated by a multi-agent pipeline. Digital Pantheon addresses the persona engineering problem of keeping agents in character under pressure, introducing an auditable multi-agent negotiation architecture. Together they map the latest solutions to three core agent challenges: runtime design, workflow orchestration, and persona engineering.

AI Agent Arxiv Digest — 2026-07-20

Three papers examining real-world challenges for AI coding agents: the first systematically demonstrates how coding agents can be tricked into supply-chain attacks via manipulated READMEs, with defenses depending more on the harness than the model; the second introduces BPO, a reinforcement learning algorithm that branches only at high-entropy decision points for more efficient agent training; the third shows how MCP can serve as a standard protocol for connecting agents to domain-specific simulation tools in industrial settings like power grids, providing a replicable template for vertical-domain agent deployment.

AI Agent Arxiv Digest — 2026-07-18

Three papers tackle production-grade agent reliability from different angles: MemCon models memory operations as an RL problem so agents learn when to store, retrieve, and forget — up to +15.2 points on 6 benchmarks; AgentCheck turns MCP servers into a debugging surface for reproducing tool faults and verifying fixes, filling a long-standing gap in the MCP ecosystem; AgentAbstain uses 263 paired tasks to show that even the strongest frontier models score below 60% on 'should-not-act' scenarios, and abstention ability barely correlates with task-solving ability — swapping in a stronger model won't fix this.

AI Agent Arxiv Digest — 2026-07-16

Three papers converge on the same question: how should each execution unit of an agent be designed so it's auditable, reusable, and recoverable at minimal blast radius when things go wrong? ATG decomposes tasks into DAGs for parallel subtask execution and intermediate result reuse; PalmClaw wraps native mobile APIs as structured tools, ditching brittle GUI click sequences; IoAT extends agent networks into the physical IoT world — from smart buildings to edge devices — sketching a coordination blueprint across cloud, edge, and sensor layers. Common thread: execution boundaries must be crisp, actions must be auditable, and failures must be locally recoverable.

AI Agent Arxiv Digest — 2026-07-07

All three papers today center on making agent systems safer, more predictable, and less failure-prone. The first two come from the same research group and take a static-analysis angle: one systematically uncovers why and how often agents get stuck in infinite loops, while the other builds dependency graphs for entire agent codebases to enable security audits and component inventories. The third targets multi-agent software development, introducing LLM confidence scores into the collaboration flow to prevent early hallucinations from cascading downstream.

AI Agent Arxiv Digest — 2026-07-01

Three papers spanning distinct dimensions of the AI Agent ecosystem: Qwen introduces the first Language World Model covering seven agent domains, enabling agents to train in simulated environments instead of relying on real APIs; Kuaishou's AgentX demonstrates industrial-scale multi-agent deployment, boosting recommendation algorithm iteration efficiency to 13.8x human output; OpenAI uses real Codex usage data to quantify how agentic AI is reshaping work across job functions, revealing that non-technical roles (legal, research) see even greater agentic dividends than engineers.

AI Agent Arxiv Digest — 2026-06-26

Three papers, three angles: **RigorBench** evaluates coding agents on process discipline rather than just pass rates, introducing five dimensions of engineering rigor; a production-focused paper shows how to customize and accelerate large multi-agent systems for enterprise use (4.48x throughput gain); and a governance paper proposes a formal protocol language for specifying human-agent boundaries in the SDLC — turning 'which decisions AI can make' from a line in a prompt into a machine-verifiable spec. Together they cover evaluation, deployment, and governance.

AI Agent Arxiv Digest — 2026-06-24

Three papers on agent platform infrastructure gaps: PlanBench-XL reveals top LLMs collapse under tool failure in large-scale ecosystems (GPT-5.4 drops from 52% to 11%); TU Munich provides the first technical taxonomy of 9 agent communication protocols (MCP/A2A/ACP/ANP) for principled selection; AMD's Arbor uses tree search as a shared cognition space for multi-agent collaboration, turning failures into useful exploration signals. Together, they outline three foundational infrastructure gaps in 2026 agent platforms.

AI Agent Arxiv Digest — 2026-06-14

Three papers tackle the same core question from different angles: **how to evaluate and operate AI Agents under real deployment conditions.** Emergence World builds a multi-agent sandbox that runs continuously for weeks, exposing behavioral drift and cross-model contamination invisible to short-term benchmarks; a survey paper establishes a complete taxonomy for agent environment design (8 attributes x 8 domains) and proposes symbolic vs. neural synthesis paradigms; Martin Monperrus's position paper declares outright that coding agents have crossed the threshold and human code review can retire.

AI Agent Arxiv Digest — 2026-06-08

Three papers mapping to three layers of the agent platform stack: AgentJet (training layer) introduces a distributed framework for simultaneous RL training of multiple heterogeneous LLMs, solving the fundamental limitation of single-model-only training tools; AdaPlanBench (evaluation layer) reveals with a 67.75% ceiling that LLM agents are far from ready for real-world scenarios where rules are disclosed progressively — it is the first benchmark to systematically quantify this adaptive planning capability; Beyond Tokens (communication layer) surveys multi-agent systems that replace text with embeddings for inter-agent communication, providing a taxonomy to evaluate the engineering trade-offs of this new communication path.

AI Agent Arxiv Digest — 2026-06-01

Three papers today zero in on the cost-capability frontier of agent deployment at scale: SR²AM redesigns planning architecture so a 30B model uses 90% fewer tokens while competing with 685B-1T systems; GroupMemBench reveals that existing memory systems completely fall apart in multi-party group conversations (the best system hits only 46% accuracy, and 1990s BM25 keyword search actually beats it); AgentFloor confirms with 16,542 test runs that the bulk of short-range tool use in agent pipelines simply doesn't need a large model. The common thread: under compute cost pressure, precisely determining 'how much intelligence each component needs' has become the central design challenge for agent platforms.

AI Agent Arxiv Digest — 2026-05-30

Three papers tackle AI Agent practice from three angles: a design language, a security map, and cognitive limitations. The first builds a two-axis classification framework giving engineers and researchers a shared vocabulary for agent architecture trade-offs; the second systematically catalogs safety and privacy risks across tool calls, memory, and multi-step execution in agentic AI; the third is the most impactful — a large-scale experiment with nearly 40,000 AI-generated ideas reveals that AI research agents tend to circle existing literature rather than genuinely broadening scientific exploration.

AI Agent Arxiv Digest — 2026-05-25

Three papers on the most pressing question for agent platforms in 2026: can safety constraints in multi-agent systems actually hold up during execution? 2605.10481 names a new failure mode — 'constraint drift': safety rules written at design time silently weaken as they pass through agent delegation, memory read/write, and tool calls, arriving at the output already distorted. 2605.07728 (SARC) proposes an architectural fix: compile regulations into four enforceable checkpoints embedded in the agent execution loop — no more relying on prompt reminders — and is open-sourced. 2605.13851 uses psychology experiments to show that when a multi-agent system's coordinator is invisible, the system's protective behaviors drop significantly — a direct design warning for mainstream orchestrator-based architectures.