id
stringlengths
8
8
section
stringclasses
11 values
subcategory
stringlengths
7
43
rtype
stringclasses
6 values
title
stringlengths
10
131
url
stringlengths
23
137
venue
stringlengths
3
74
year
int64
1.98k
2.03k
authors
stringlengths
4
254
description
stringlengths
99
269
why
stringlengths
77
200
evidence
stringclasses
7 values
layer
stringclasses
9 values
age-0102
Benchmarks & Datasets
Prompt-injection evaluation
Benchmark
AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents
https://openreview.net/forum?id=m1YYAQjO3w
NeurIPS Datasets & Benchmarks
2,024
Edoardo Debenedetti; Jie Zhang; Mislav Balunovic; Luca Beurer-Kellner; Marc Fischer; Florian Tramèr
Provides an extensible adversarial environment with 97 realistic tasks and 629 security test cases for agents executing tools over untrusted data.
Tests the instruction-data boundary that every external-input and artifact edge must preserve; compromised peer-agent messages require separate evaluation.
Benchmark/dataset
Reliability
age-0103
Benchmarks & Datasets
Agent misuse benchmark
Benchmark
AgentHarm: A Benchmark for Measuring Harmfulness of LLM Agents
https://openreview.net/forum?id=AC5n7xHuR1
ICLR
2,025
Maksym Andriushchenko; Alexandra Souly; Mateusz Dziemian; Derek Duenas; Maxwell Lin; Justin Wang; Dan Hendrycks; Andy Zou; J. Zico Kolter; Matt Fredrikson; Yarin Gal; Xander Davies
Evaluates refusal and retained task capability on 110 explicitly malicious multi-stage agent tasks with 440 augmented variants across 11 harm categories.
Tests whether safety gates reject prohibited work and stop a compromised node from completing a harmful path; the score is not a general measure of agent-system safety.
Benchmark/dataset
Gates
age-0104
Reliability & Durable Execution
Structural prompt-injection defense
Paper
IPIGuard: A Novel Tool Dependency Graph-Based Defense Against Indirect Prompt Injection in LLM Agents
https://aclanthology.org/2025.emnlp-main.53/
EMNLP
2,025
Hengyu An; Jinghuai Zhang; Tianyu Du; Chunyi Zhou; Qingming Li; Tao Lin; Shouling Ji
Models task execution as traversal over a planned Tool Dependency Graph and separates action planning from interaction with untrusted external data.
Shows how structural constraints on allowed tool edges can block unauthorized actions instead of relying only on prompts or classifiers.
Peer-reviewed research
Work graphs
age-0105
Protocols & Handoffs
Agent identity and delegated authorization
Docs
Identity Management for Agentic AI: The new frontier of authorization, authentication, and security for an AI agent world
https://openid.net/wp-content/uploads/2025/10/Identity-Management-for-Agentic-AI.pdf
OpenID Foundation
2,025
OpenID Foundation Artificial Intelligence Identity Management Community Group
Maps identity, authentication, delegated authority, consent, governance, and accountability gaps for autonomous agents against OAuth and OpenID Connect foundations.
Provides a standards-grounded vocabulary for identity-bearing nodes and auditable delegation across agent and tool edges; it is a non-normative landscape whitepaper, not a protocol.
Official documentation
Handoffs
age-0106
Protocols & Handoffs
Agent capability and discovery schema
Standard
Open Agentic Schema Framework (OASF) v1.1.0
https://github.com/agntcy/oasf/releases/tag/v1.1.0
AGNTCY Project, Linux Foundation
2,026
AGNTCY Project
Defines versioned schemas and taxonomies for agent identity metadata, capabilities, skills, modules, interactions, and discovery records, including MCP and A2A alignment modules.
Gives graph builders a portable node descriptor for capability-based discovery and routing validation; it describes capabilities but does not execute a graph or authorize an edge.
Industry standard
Roles
age-0107
State, Memory & Artifacts
Artifact provenance
Standard
PROV-DM: The PROV Data Model
https://www.w3.org/TR/prov-dm/
W3C Recommendation
2,013
Luc Moreau; Paolo Missier
Defines entities, activities, agents, derivation, attribution, responsibility, bundles, and collections for interoperable provenance records.
Maps artifacts, execution steps, and responsible nodes into a portable lineage model; storage, access control, and cryptographic integrity remain implementation concerns.
Industry standard
State
age-0108
State, Memory & Artifacts
Cryptographic artifact provenance
Paper
in-toto: Providing farm-to-table guarantees for bits and bytes
https://www.usenix.org/conference/usenixsecurity19/presentation/torres-arias
USENIX Security
2,019
Santiago Torres-Arias; Hammad Afzali; Trishank Karthik Kuppusamy; Reza Curtmola; Justin Cappos
Introduces signed material and product links for supply-chain steps and verifies them against an expected layout, supported by analysis of 30 historical compromises and deployed integrations.
Supplies a transferable receipt pattern that binds each artifact transformation to an authorized step and identity; it verifies process integrity, not semantic correctness.
Peer-reviewed research
State
age-0109
Reliability & Durable Execution
Portable workflow semantics
Standard
Serverless Workflow Specification v1.0.0
https://open-workflow-specification.org/blog/releases/release-100/
Open Workflow Specification, Cloud Native Computing Foundation
2,025
Open Workflow Specification Project
Defines a vendor-neutral workflow language with sequential and concurrent tasks, event correlation, service calls, error handling, retries, and timeouts.
Provides an inspectable execution contract for production work graphs; conformance does not itself supply a durable runtime, an agent protocol, or model-level correctness.
Industry standard
Reliability
age-0110
Critiques & Limits
Compound-system call scaling
Paper
Are More LLM Calls All You Need? Towards the Scaling Properties of Compound AI Systems
https://proceedings.neurips.cc/paper_files/paper/2024/hash/51173cf34c5faac9796a47dc2fdd3a71-Abstract-Conference.html
NeurIPS
2,024
Lingjiao Chen; Jared Davis; Boris Hanin; Peter Bailis; Ion Stoica; Matei Zaharia; James Zou
Analyzes Vote and Filter-Vote compound systems and finds that performance can rise and then fall as the number of language-model calls increases because query difficulty is heterogeneous.
Bounds claims that adding calls, voters, or nodes automatically improves a system; the experiments cover simple aggregation systems rather than rich agent organizations.
Peer-reviewed research
Observability & cost
age-0111
Verification & Evals
Verification-aware work graphs
Paper
Verification-Aware Planning for Multi-Agent Systems
https://aclanthology.org/2026.eacl-long.353/
EACL
2,026
Tianyang Xu; Dan Zhang; Kushan Mitra; Estevam Hruschka
Introduces VeriMAP, which decomposes tasks into a dependency graph and attaches planner-defined Python and natural-language verification functions to subtasks before execution.
Makes acceptance criteria part of the work graph so handoff failures can trigger local repair instead of surfacing only in a final answer.
Peer-reviewed research
Gates
age-0112
Research Foundations
Human-in-the-loop work-graph planning
Paper
AIPOM: Agent-aware Interactive Planning for Multi-Agent Systems
https://aclanthology.org/2025.emnlp-demos.7/
EMNLP System Demonstrations
2,025
Hannah Kim; Kushan Mitra; Chen Shen; Dan Zhang; Estevam Hruschka
Presents conversational and graph-based interfaces for inspecting, editing, and collaboratively guiding plans in orchestrated multi-agent systems.
Treats the work graph as a human-legible control surface; its non-autonomous agents execute assigned tasks, so it is a planning substrate rather than a complete agent organization.
Peer-reviewed research
Work graphs
age-0113
Frameworks & SDKs
Graph workflow runtime
Blog
Build reliable multi-agent applications with ADK Go 2.0
https://developers.googleblog.com/announcing-adk-go-20/
Google Developers Blog
2,026
Toni Klopfenstein; Sampath Kumar Maddula
Documents a graph-based multi-agent workflow engine with typed nodes, conditional edges, fan-out and fan-in, nested graphs, cycles, retries, durable human input, state, and unified telemetry.
Provides a first-party implementation of explicit, resumable work graphs while showing that function and tool nodes remain distinct from agent nodes.
Practitioner analysis
Work graphs
age-0114
Research Foundations
Decentralized adaptive coordination
Paper
AgentNet: Decentralized Evolutionary Coordination for LLM-based Multi-Agent Systems
https://proceedings.neurips.cc/paper_files/paper/2025/hash/9a379c1b05793d1c42dc832269834515-Abstract-Conference.html
NeurIPS
2,025
Yingxuan Yang; Huacan Chai; Shuai Shao; Yuanyi Song; Siyuan Qi; Renting Rui; Weinan Zhang
Organizes specialized agents in a decentralized DAG, uses retrieval-backed local experience for capability refinement, and adapts routing without a single central orchestrator.
Offers a contrasting topology for settings where central coordination is a bottleneck or trust boundary; its reported evidence is concentrated in reasoning and coding tasks.
Peer-reviewed research
Topology
age-0115
Reliability & Durable Execution
Effect-typed agent contracts
Paper
ETAS: An Effect-Typed Language for Agent Systems
https://arxiv.org/abs/2607.17780
arXiv
2,026
Huiri Tan; Yikun Wang; Puyang Zhang; Shangyu Li; Jiasi Shen
Defines a language in which agents, tools, typed memory, approvals, policies, effects, and execution traces are semantic program elements, with static obligations and runtime monitors.
Provides a transferable contract model for authorizing and auditing node actions before and during execution; it is a new preprint and not yet deployment evidence.
Research preprint
Reliability
age-0116
Critiques & Limits
Consensus-induced search collapse
Paper
The Shared Discovery Paradox: How a One-Answer Rule Turns Better Information into Worse Search
https://arxiv.org/abs/2607.18045
arXiv
2,026
Yohei Nakajima
Uses an exactly solvable multi-searcher benchmark to show that pooling information into one repeated recommendation can improve belief accuracy while sharply reducing collective discovery coverage.
Separates information quality from allocation policy and warns that consensus edges can collapse parallel exploration unless a coordinator preserves a portfolio of actions.
Research preprint
Topology
age-0117
Critiques & Limits
Relay information bottlenecks
Paper
When Do Multi-Agent Systems Help? An Information Bottleneck Perspective
https://arxiv.org/abs/2607.16133
arXiv
2,026
Wendi Yu; Lianhao Zhou; Xiangjue Dong; Sai Sudarshan Barath; Declan Staunton; Byung-Jun Yoon; Xiaoning Qian; James Caverlee; Shuiwang Ji
Formalizes bounded inter-agent relays as an information bottleneck and reports 18 controlled experiments across five benchmarks and three model scales.
Explains when isolated contexts and compressed handoffs save useful context and when they discard task-relevant information, providing a direct test for whether a graph earns its edge losses.
Research preprint
Handoffs
age-0118
Critiques & Limits
Critique uptake failure
Paper
Precise but Uncoupled: Reviewer Precision Does Not Guarantee Critique Uptake in Multi-Agent Math Reasoning
https://arxiv.org/abs/2607.15388
arXiv
2,026
Chih-Hsuan Yang; Jingyan Jiang; Vikram Vasudevan; Cheng-Hau Yang; Huihuo Zheng; Le Chen; Eliu A. Huerta; Venkatram Vishwanath; Ian T. Foster; Rajeev Thakur
Evaluates 4,181 verifier-grounded math problems and finds that a precise reviewer can still yield weak repair when the protocol does not carry useful critique into the next candidate.
Shows that a verifier node is insufficient by itself: the outgoing edge must couple accepted evidence to a concrete state change or retry action.
Research preprint
Gates
age-0119
Research Foundations
Grammar-based system search
Paper
Grammar Search for Multi-Agent Systems
https://aclanthology.org/2026.acl-long.75/
ACL
2,026
Mayank Singh; Vikas Yadav; Shiva Krishna Reddy Malay; Shravan Nayak; Sai Rajeswar; Sathwik Tejaswi Madhusudhan; Eduardo Blanco
Searches a fixed grammar of composable multi-agent components to construct modular, interpretable organizations at lower search cost than unconstrained code generation.
Makes the agent-system design space explicit and auditable while retaining automated search over roles, operators, and coordination structure.
Peer-reviewed research
Evolution
age-0120
Research Foundations
Topology policy optimization
Paper
Graph-GRPO: Stabilizing Multi-Agent Topology Learning via Group Relative Policy Optimization
https://aclanthology.org/2026.findings-acl.1010/
Findings of ACL
2,026
Yueyang Cang; Xiaoteng Zhang; Erlu Zhao; Zehua Ji; Yuhang Liu; Yuchen He; Zhiyuan Ning; Chen Yijun; Wenge Que; Li Shi
Samples groups of query-specific communication graphs and uses relative rewards to assign edge-level credit while training a topology-generating policy.
Addresses unstable topology learning with explicit graph samples and localized credit rather than treating orchestration as an opaque prompt update.
Peer-reviewed research
Evolution
age-0121
Observability & Cost
Cost-aware graph compression
Paper
AgentSlimming: Towards Efficient and Cost-Aware Multi-Agent Systems
https://aclanthology.org/2026.acl-long.1387/
ACL
2,026
Yulang Chen; Haoxuan Peng; Jinyan Liu; Zichen Wen; Dongrui Liu; Linfeng Zhang
Compresses graph-structured agent workflows using centrality and approximate Shapley importance, then validates agent removal and cheaper-model substitution against the original system.
Provides a measured way to reduce nodes and spend without assuming every role contributes equally or that a smaller graph preserves behavior automatically.
Peer-reviewed research
Evolution
age-0122
Research Foundations
One-shot topology generation
Paper
TopoDIM: One-shot Topology Generation of Diverse Interaction Modes for Multi-Agent Systems
https://aclanthology.org/2026.findings-acl.207/
Findings of ACL
2,026
Rui Sun; Jie Ding; Chenghua Gong; Tianjun Gu; Yihang Jiang; Juyuan Zhang; Liming Pan; Linyuan Lü
Generates heterogeneous decentralized communication topologies in one pass, replacing repeated coordination rounds with a task-conditioned interaction structure.
Expands topology design beyond homogeneous or fully connected teams while directly evaluating the quality and token cost of the generated graph.
Peer-reviewed research
Topology
age-0123
Research Foundations
Evolving hypergraph collaboration
Paper
EvoHyper: Evolving Hypergraph Topologies for Unified Collaboration in Multi-Agent Communication
https://aclanthology.org/2026.findings-acl.1258/
Findings of ACL
2,026
Heng Zhang; Yihao Zhong; Lubin Gan; Zhihe Chen; Jiajun Wu; Yuling Shi; Xiaodong Gu; Hao Zhang; Haochen You; Jin Huang
Uses evolving hyperedges to bind groups of agents to shared memory while a controller creates, updates, and merges collaboration units during execution.
Models group communication as a first-class structure and joins topology evolution with explicit shared-state ownership instead of reducing every exchange to pairwise chat.
Peer-reviewed research
Evolution
age-0124
Research Foundations
Heterogeneous graph routing
Paper
RouterHGC: Optimized Router for LLM-based Multi-Agent Systems via Heterogeneous Graph Contrastive Learning
https://aclanthology.org/2026.findings-acl.1589/
Findings of ACL
2,026
Yitao Xiao; Shaoyong Guo; Guoming Yang; Qingnan Wang; Yinlin Ren; Xuesong Qiu; Qi Feng
Represents queries, collaboration modes, roles, and language models as heterogeneous nodes and learns cost-performance routing relationships with graph contrastive learning.
Treats the agent organization itself as the graph under optimization and exposes model, role, mode, and budget choices within one routing decision.
Peer-reviewed research
Evolution
age-0125
Research Foundations
Dynamic workflow scheduling
Paper
LLM-as-Scheduler: Agentic Workflow Dynamic Scheduling
https://aclanthology.org/2026.acl-long.581/
ACL
2,026
Dawei Xiang; Kexin Chu; Wenyan Xu; Wenhui Zhang; Wei Zhang
Dynamically selects query-specific paths through a workflow DAG, including early exit, verification, repair, and rerouting decisions made during execution.
Separates the reusable work graph from the runtime path and supplies concrete control points for budgets, validation, and local recovery.
Peer-reviewed research
Work graphs
age-0126
Research Foundations
Adaptive message representation
Paper
Learning Optimal Message Representations for Agentic Communication
https://aclanthology.org/2026.findings-acl.1441/
Findings of ACL
2,026
Shashwat Gupta; Anson Bastos; Mayukh Das; Supriyo Ghosh; Nagarajan Natarajan; Chetan Bansal; Saravan Rajmohan
Models the choice of inter-agent message representation as an expanding Markov decision process that can select natural-language or structured protocols at each exchange.
Makes payload format an adaptive edge policy rather than a global convention, clarifying when structure, compression, or expressiveness should dominate.
Peer-reviewed research
Handoffs
age-0127
Reliability & Durable Execution
Handoff clarification
Paper
AgentAsk: Multi-Agent Systems Need to Ask
https://aclanthology.org/2026.acl-long.1294/
ACL
2,026
Bohan Lin; Kuo Yang; Zelin Tan; Yingchuan Lai; Chen Zhang; Guibin Zhang; Xinlei Yu; Miao Yu; Xu Wang; Yudong Zhang; Yang Wang
Defines an edge-level handoff-error taxonomy and adds selective clarification questions at critical exchanges where transferred information is ambiguous or incomplete.
Turns clarification into an explicit recovery action for communication-contract failures instead of letting uncertainty silently propagate downstream.
Peer-reviewed research
Handoffs
age-0128
Reliability & Durable Execution
Execution-verified diagnosis
Paper
Towards Self-Improving Error Diagnosis in Multi-Agent Systems
https://aclanthology.org/2026.findings-acl.98/
Findings of ACL
2,026
Jiazheng Li; Emine Yilmaz; Bei Chen; Thu Le
Detects local anomalies, traces backward to decisive failures, tests diagnostic hypotheses with tools, and retains only execution-verified diagnostic memories.
Combines trace localization, active validation, and durable learning so a diagnostic node improves from evidence rather than storing unverified explanations.
Peer-reviewed research
Observability & cost
age-0129
Reliability & Durable Execution
Denial-of-collaboration attacks
Paper
CORBA: Contagious Recursive Blocking Attacks on Multi-Agent Systems Based on Large Language Models
https://aclanthology.org/2026.findings-acl.342/
Findings of ACL
2,026
Zhenhong Zhou; Zherui Li; Jie Zhang; Yuanhe Zhang; Kun Wang; Yang Liu; Qing Guo
Formalizes attacks that deform communication topology into cycles or recursively expanding exchanges, blocking useful collaboration while exhausting system resources.
Shows that edge activation and termination policies are security boundaries, not only scheduling details, and motivates cycle and budget guards.
Peer-reviewed research
Reliability
age-0130
Reliability & Durable Execution
Topology confidentiality
Paper
CIA: Inferring the Communication Topology from LLM-based Multi-Agent Systems
https://aclanthology.org/2026.acl-long.815/
ACL
2,026
Yongxuan Wu; Xixun Lin; He Zhang; Nan Sun; Kun Wang; Chuan Zhou; Shirui Pan; Yanan Cao
Infers hidden communication topology from black-box multi-agent systems using adversarial queries and semantic correlations among observed responses.
Establishes topology confidentiality as an engineering property and demonstrates that apparently private coordination structure can leak through behavior.
Peer-reviewed research
Reliability
age-0131
Benchmarks & Datasets
Cascading-injection evaluation
Benchmark
ACIArena: Toward Unified Evaluation for Agent Cascading Injection
https://aclanthology.org/2026.acl-long.457/
ACL
2,026
Hengyu An; Minxi Li; Jinghuai Zhang; Naen Xu; Chunyi Zhou; Changjiang Li; Xiaogang Xu; Tianyu Du; Shouling Ji
Evaluates cascading injection through external inputs, agent profiles, and inter-agent messages across six implementations and 1,356 attack cases.
Measures how topology, roles, and interaction design affect propagation instead of treating prompt injection as an isolated single-node event.
Benchmark/dataset
Reliability
age-0132
Reliability & Durable Execution
Unknown-attack detection
Paper
BlindGuard: Safeguarding LLM-based Multi-Agent Systems under Unknown Attacks
https://aclanthology.org/2026.acl-long.1819/
ACL
2,026
Rui Miao; Yixin Liu; Yili Wang; Xu Shen; Yue Tan; Yiwei Dai; Shirui Pan; Xin Wang
Learns normal individual, neighborhood, and global interaction behavior without attack labels, then detects malicious agents through corruption-guided contrastive learning.
Adds graph-aware anomaly detection for attacks that were absent from training and distinguishes local behavior from neighborhood and system-wide patterns.
Peer-reviewed research
Reliability
age-0133
Reliability & Durable Execution
Fine-grained graph anomaly detection
Paper
Explainable and Fine-Grained Safeguarding of LLM Multi-Agent Systems via Bi-Level Graph Anomaly Detection
https://aclanthology.org/2026.acl-long.1407/
ACL
2,026
Junjun Pan; Yixin Liu; Rui Miao; Kaize Ding; Yu Zheng; Quoc Viet Hung Nguyen; Alan Wee-Chung Liew; Shirui Pan
Detects sentence- and token-level anomalies over multi-agent interaction graphs and attributes suspicious tokens to support finer-grained explanations of compromised exchanges.
Connects graph-level protection to actionable message evidence rather than returning only an opaque risk score for an entire agent or run.
Peer-reviewed research
Reliability
age-0134
Verification & Evals
Collective-intelligence prediction
Paper
Identifying Collective Intelligence Factor in LLM Agent Groups for Generalizable Multi-Agent System Design
https://aclanthology.org/2026.findings-acl.624/
Findings of ACL
2,026
Zhilun Zhou; Zihan Liu; Jiahe Liu; Yihan Wang; Qingyu Shao; Fengli Xu; Depeng Jin; Yong Li
Studies 108 agent groups varying team size, model composition, and communication topology, then derives a factor intended to predict transferable system designs.
Moves evaluation beyond one benchmark score by testing whether measurable group properties forecast performance across designs and tasks.
Peer-reviewed research
Evolution
age-0135
Critiques & Limits
Communication capacity bounds
Paper
Benefits and Limitations of Communication in Multi-Agent Reasoning
https://openreview.net/forum?id=0aPIVJUz5T
ICLR
2,026
Michael Rizvi-Martel; Satwik Bhattamishra; Neil Rathi; Guillaume Rabusseau; Michael Hahn
Derives bounds connecting agent count and the quantity and structure of communication to exact reasoning and speedups, then tests the tradeoffs on controlled language-model tasks.
Supplies principled limits for node count and bandwidth decisions and identifies regimes where communication helps or becomes an intrinsic bottleneck.
Peer-reviewed research
Handoffs
age-0136
Reliability & Durable Execution
Topology-aware prompt attacks
Paper
Agents Under Siege: Breaking Pragmatic Multi-Agent LLM Systems with Optimized Prompt Attacks
https://aclanthology.org/2025.acl-long.476/
ACL
2,025
Rana Shahroz; Zhen Tan; Sukwon Yun; Charles Fleming; Tianlong Chen
Optimizes attack placement across latency- and bandwidth-constrained agent networks using graph-flow formulations to disrupt multi-agent task completion.
Makes topology part of the adversary model and tests how limited attack budgets exploit particular communication paths and bottlenecks.
Peer-reviewed research
Reliability
age-0137
Benchmarks & Datasets
Networked-agent coordination
Paper
AgentsNet: Coordination and Collaborative Reasoning in Multi-Agent LLMs
https://arxiv.org/abs/2507.08616
arXiv; ICML 2025 Workshop on Collaborative and Federated Agentic Workflows
2,025
Florian Grötschla; Luis Müller; Jan Tönshoff; Mikhail Galkin; Bryan Perozzi
Evaluates local messaging and learned coordination protocols on distributed graph problems over explicit network topologies, with experiments scaling to 100 agents.
Provides unusually direct evidence about decentralized protocol formation and topology-aware reasoning; the selected artifact remains a preprint and workshop contribution.
Research preprint
Topology
age-0138
Benchmarks & Datasets
Pure coordination games
Benchmark
LLM-Coordination: Evaluating and Analyzing Multi-agent Coordination Abilities in Large Language Models
https://aclanthology.org/2025.findings-naacl.448/
Findings of NAACL
2,025
Saaket Agashe; Yue Fan; Anthony Reyna; Xin Eric Wang
Evaluates agentic play in four pure-coordination games and 198 questions covering environment comprehension, theory of mind, and joint planning.
Separates coordination capability into measurable components and partner-generalization tests, although communication topology is not its central experimental variable.
Benchmark/dataset
Gates
age-0139
Research Foundations
Relevance-ordered message passing
Paper
Graph-of-Agents: A Graph-based Framework for Multi-Agent LLM Collaboration
https://openreview.net/forum?id=34cANdsHKV
ICLR
2,026
Sukwon Yun; Jie Peng; Pingzhi Li; Wendong Fan; Jie Chen; James Y. Zou; Guohao Li; Tianlong Chen
Selects relevant agents from model cards, constructs directed edges from response relevance, performs forward and reverse message passing, and pools the revised answers.
Offers an end-to-end graph construction and aggregation method in which node selection, edge order, and communication direction are explicit design choices.
Peer-reviewed research
Topology
age-0140
Research Foundations
Response-conditioned self-organization
Paper
Stochastic Self-Organization in Multi-Agent Systems
https://openreview.net/forum?id=rS3Jb9AAej
ICLR
2,026
Nurbek Tastan; Samuel Horváth; Karthik Nandakumar
Builds a response-conditioned directed acyclic graph from approximate peer-contribution scores and updates the communication structure across collaboration rounds.
Adapts topology from observed agent behavior without a separate judge, pretrained generator, or fixed graph, while retaining stable acyclic message flow.
Peer-reviewed research
Evolution
age-0141
Research Foundations
Memory-augmented agent routing
Paper
GraphPlanner: Graph Memory-Augmented Agentic Routing for Multi-Agent LLMs
https://openreview.net/forum?id=ZdGB7MNQDT
ICLR
2,026
Tao Feng; Haozhen Zhang; Zijie Lei; Peixuan Han; Jiaxuan You
Generates query-specific workflows by selecting both language-model backbones and agent roles while consulting a heterogeneous graph of query-agent-response history.
Combines work-graph routing with reusable interaction memory and evaluates unseen-task and unseen-model generalization rather than only in-distribution quality.
Peer-reviewed research
Evolution
age-0142
Observability & Cost
Runtime adaptive supervision
Paper
Stop Wasting Your Tokens: Towards Efficient Runtime Multi-Agent Systems
https://openreview.net/forum?id=pzFhtpkabh
ICLR
2,026
Fulin Lin; Shaowen Chen; Ruishan Fang; Hongwei Wang; Tao Lin
Adds a lightweight supervisor triggered by an LLM-free filter to intervene when a multi-agent run shows errors, inefficient behavior, or contaminated observations.
Places cost and failure control inside the live execution path and tests whether selective supervision can reduce token use without redesigning the base graph.
Peer-reviewed research
Observability & cost
age-0143
Benchmarks & Datasets
Multi-agent error attribution
Benchmark
Aegis: Automated Error Generation and Attribution for Multi-Agent Systems
https://openreview.net/forum?id=zqcYoxXiN3
ICLR
2,026
Fanqi Kong; Ruijie Zhang; Huaxiao Yin; Guibin Zhang; Xiaofei Zhang; Ziang Chen; Zhaowei Zhang; Xiaoyuan Zhang; Song-Chun Zhu; Xue Feng
Creates 9,533 trajectories with annotated faulty agents and error modes by injecting context-aware faults into successful runs across multiple architectures and domains.
Provides graph-component labels for training and evaluating failure attribution instead of relying only on final-task success or expensive manual trace review.
Benchmark/dataset
Reliability
age-0144
Frameworks & SDKs
Multi-agent reinforced training
Paper
MARTI: A Framework for Multi-Agent LLM Systems Reinforced Training and Inference
https://openreview.net/forum?id=E7jZqo0A50
ICLR
2,026
Kaiyan Zhang; Kai Tian; Runze Liu; Sihang Zeng; Xuekai Zhu; Guoli Jia; Yuchen Fan; Xingtai Lv; Yuxin Zuo; Che Jiang; Yuru Wang; Jianyu Wang; Ermo Hua; Xinwei Long; Junqi Gao; Youbang Sun; Zhiyuan Ma; Ganqu Cui; Ning Ding; Biqing Qi; Bowen Zhou
Supports centralized multi-agent interaction, distributed policy training, asynchronous rollouts, graph-defined workflows, heterogeneous models, tools, and configurable rewards.
Connects executable organizations to reinforced training infrastructure so graph policies can be optimized and evaluated rather than remaining hand-authored inference code.
Peer-reviewed research
Evolution
age-0145
Research Foundations
Parallel message propagation
Paper
MPAS: Breaking Sequential Constraints of Multi-Agent Communication Topologies via Individual-Epistemic Message Propagation
https://ojs.aaai.org/index.php/AAAI/article/view/40231
AAAI
2,026
Jingxuan Yu; Ju Jia; Simeng Qin; Xiaojun Jia; Siqi Ma; Yihao Huang; Yali Yuan; Guang Cheng
Reworks sequential multi-agent communication into parallel graph message propagation with optimized connections and aggregation while preserving agent-specific epistemic state.
Targets latency introduced by dependency-ordered exchanges and makes the tradeoff between parallel execution, aggregation, and resilience measurable.
Peer-reviewed research
Topology
age-0146
Verification & Evals
Trajectory failure diagnosis
Benchmark
AgentRx: Diagnosing AI Agent Failures from Execution Trajectories
https://arxiv.org/abs/2602.02475
arXiv
2,026
Shraddha Barke; Arnav Goyal; Alind Khare; Avaljot Singh; Suman Nath; Chetan Bansal
Releases 115 manually annotated failed trajectories and diagnoses critical failure steps by synthesizing guarded executable constraints and evidence-bearing validation logs.
Provides a concrete bridge from heterogeneous agent traces to repeatable failure localization; its evaluation includes Magentic-One but also single-agent workflows.
Benchmark/dataset
Gates
age-0147
Verification & Evals
Cross-framework system evaluation
Benchmark
MAESTRO: Multi-Agent Evaluation Suite for Testing, Reliability, and Observability
https://arxiv.org/abs/2601.00481
arXiv
2,026
Tie Ma; Yixi Chen; Vaastav Anand; Alessandro Cornacchia; Amândio R. Faustino; Guanheng Liu; Shan Zhang; Hongbin Luo; Suhaib A. Fahmy; Zafar A. Qazi; Marco Canini
Standardizes configuration and execution for 12 representative multi-agent systems and exports framework-agnostic traces with latency, cost, reliability, and failure signals.
Enables controlled comparison across architectures and repeated runs while revealing that topology can dominate model or tool changes in resource and reliability profiles.
Benchmark/dataset
Observability & cost
age-0148
Observability & Cost
Causal root-cause tracing
Paper
AgentTrace: Causal Graph Tracing for Root Cause Analysis in Deployed Multi-Agent Systems
https://arxiv.org/abs/2603.14688
arXiv; ICLR 2026 Workshop on Agents in the Wild
2,026
Zhaohui Geoffrey Wang
Reconstructs causal graphs from execution logs, traces backward from observed failures, and ranks candidate root causes without requiring language-model inference during debugging.
Offers an interpretable, low-latency diagnosis primitive for cascading failures while remaining a preprint and workshop result rather than broad production validation.
Research preprint
Observability & cost
age-0149
Start Here
Production reference architecture
Docs
Multi-Agent Reference Architecture
https://github.com/microsoft/multi-agent-reference-architecture
Microsoft
2,026
Microsoft
Maps agent registries, communication, memory, observability, evaluation, security, governance, and deployment into an end-to-end reference architecture for multi-agent systems.
Provides a broad engineering checklist grounded in customer implementations; it is evolving guidance in an official repository, not a versioned standard.
Official documentation
Topology
age-0150
Frameworks & SDKs
Event-driven agent runtime
Docs
AutoGen Core
https://microsoft.github.io/autogen/stable/user-guide/core-user-guide/index.html
Microsoft AutoGen Documentation
2,026
Microsoft
Documents an actor-style agent runtime with asynchronous messages, topics, subscriptions, distributed execution, resilience mechanisms, and telemetry.
Adds the lower-level message and runtime architecture beneath conversational teams, clarifying how agent nodes communicate and scale across processes.
Official documentation
Handoffs
age-0151
Frameworks & SDKs
Managed multi-agent orchestration
Docs
Multiagent orchestration
https://platform.claude.com/docs/en/managed-agents/multiagent-orchestration
Claude Platform Documentation
2,026
Anthropic
Documents managed multi-agent orchestration with a lead agent that creates specialized subagents, delegates work, and combines their results under a beta API contract.
Provides a current first-party implementation of dynamic specialization and hierarchical delegation while clearly exposing its managed-runtime boundary and beta status.
Official documentation
Topology
age-0152
Reliability & Durable Execution
Distributed agent runtime
Blog
Introducing Agent Executor, Google’s distributed Agent Runtime
https://cloud.google.com/blog/products/ai-machine-learning/agent-executor-googles-distributed-agent-runtime
Google Cloud Blog
2,026
Jaana Dogan; Ethan Bao
Introduces an open-source runtime standard for execution, resumption, isolation, session consistency, reconnection, trajectory branching, and federated agent deployment.
Addresses durable long-running graphs at the runtime layer and makes checkpoints, single-writer state, recovery, and deployment boundaries explicit; the release is in preview.
Practitioner analysis
Reliability
age-0153
Observability & Cost
Workflow profiling and optimization
Tool
NVIDIA NeMo Agent Toolkit
https://github.com/NVIDIA/NeMo-Agent-Toolkit
NVIDIA
2,026
NVIDIA
Provides framework-agnostic profiling, evaluation, optimization, tracing, and A2A team support for agent workflows, including experimental graph-performance primitives.
Lets engineers measure whole-graph latency and cost while experimenting with parallel branches, speculative execution, and node priorities in maintained open-source tooling.
Maintained OSS project
Observability & cost
age-0154
Verification & Evals
Protocol conformance
Tool
A2A Protocol Technology Compatibility Kit
https://github.com/a2aproject/a2a-tck
Agent2Agent Project, Linux Foundation
2,026
Agent2Agent Project
Tests Agent2Agent protocol implementations across gRPC, JSON-RPC, and HTTP with JSON and emits machine-readable, HTML, and JUnit conformance reports.
Turns interoperability claims at agent handoff edges into executable gates that can run in development and continuous integration.
Maintained OSS project
Gates
age-0155
Protocols & Handoffs
Distributed agent discovery
Tool
AGNTCY Dir
https://github.com/agntcy/dir
AGNTCY Project, Linux Foundation
2,026
AGNTCY Project
Implements a distributed directory for capability-based discovery of agents and multi-agent systems using OASF records and cryptographic identity.
Covers how potential graph participants are registered, discovered, authenticated, and selected before a work or communication edge is created.
Maintained OSS project
Handoffs
age-0156
Reliability & Durable Execution
Runtime governance
Tool
Agent Governance Toolkit
https://github.com/microsoft/agent-governance-toolkit
Microsoft
2,026
Microsoft
Enforces runtime policies on tool calls and inter-agent messages with identity, delegation, audit logs, telemetry, and adapters for multiple agent frameworks.
Implements authorization and audit controls at graph edges; it is a public preview, and its in-process enforcement layer is not an operating-system security boundary.
Maintained OSS project
Gates
age-0157
Reliability & Durable Execution
Multi-agent threat modeling
Docs
Multi-Agentic System Threat Modeling Guide v1.0
https://genai.owasp.org/resource/multi-agentic-system-threat-modeling-guide-v1-0/
OWASP GenAI Security Project
2,025
OWASP GenAI Security Project
Applies an agentic threat taxonomy to multi-agent architectures and examines attack surfaces introduced by delegation, messaging, shared memory, tools, and cross-agent actions.
Provides a structured way to identify graph-specific trust boundaries and abuse paths; it is community security guidance rather than a certification or formal standard.
Official documentation
Gates
age-0158
Observability & Cost
Agent operations reference architecture
Blog
AgentOps: Operationalize agentic AI at scale with Amazon Bedrock AgentCore
https://aws.amazon.com/blogs/machine-learning/agentops-operationalize-agentic-ai-at-scale-with-amazon-bedrock-agentcore/
AWS Machine Learning Blog
2,026
Anastasia Tzeveleka; Anna Grüebler; Antonio Rodriguez; Sergio Garcés Vitale; Aris Tsakpinis
Organizes production agent operations around governance, build and release, multi-level evaluation, observability, identity propagation, and cost attribution in a reference architecture.
Adds concrete system measures such as orchestration accuracy and exchange quality while keeping the vendor-specific implementation context visible.
Practitioner analysis
Observability & cost
age-0159
Production Case Studies
Enterprise supervisor architecture
Blog
Powering agentic AI sales strategy with Amazon Bedrock AgentCore
https://aws.amazon.com/blogs/machine-learning/powering-agentic-ai-sales-strategy-with-amazon-bedrock-agentcore/
AWS Machine Learning Blog
2,026
Nicolle Belaunde; Umesh Mohan; Xinrui Nie
Describes an internal production system with more than 20 domain agents, supervisor routing, OAuth propagation, MCP tools, approvals, memory, distributed tracing, and evaluation.
Offers a rare detailed account of an enterprise supervisor graph at operating scale; its usage, latency, and outcome figures are self-reported by AWS.
Practitioner analysis
Topology
age-0160
Production Case Studies
Verified coding swarm
Blog
Designing Autonomous AI Agents: How We Built a Multi-Agent Swarm for UI Development
https://www.epam.com/insights/ai/blogs/building-multi-swarm-autonomous-ai-agent
EPAM Insights
2,026
Stanislau Shandrokha; Pavel Golub; Andrey Voroshkov
Describes a seven-agent state-machine workflow with resumable state, explicit roles, adversarial visual and code reviewers, and verification gates for interface development.
Reports practical failures such as overlapping work, skipped steps, and rubber-stamping alongside mitigations; the evidence is a narrow, self-reported implementation account.
Practitioner analysis
Gates
age-0161
Protocols & Handoffs
Agent message envelopes
Standard
FIPA ACL Message Structure Specification
https://www.fipa.org/specs/fipa00061/SC00061G.html
Foundation for Intelligent Physical Agents
2,002
FIPA TC Communication; FIPA Architecture Board
Defines interoperable agent messages with performatives, sender and receiver, ontology, protocol, conversation identifiers, reply correlation, and deadlines.
Supplies the historical standards basis for message-envelope and conversation concepts that reappear in current agent handoff protocols.
Industry standard
Handoffs
age-0162
Production Case Studies
Network-level red teaming
Blog
Red-teaming a network of agents: Understanding what breaks when AI agents interact at scale
https://www.microsoft.com/en-us/research/blog/red-teaming-a-network-of-agents-understanding-what-breaks-when-ai-agents-interact-at-scale/
Microsoft Research Blog
2,026
Gagan Bansal; Shujaat Mirza; Keegan Hines; Will Epperson; Zachary Huang; Whitney Maxwell; Pete Bryan; Tyler Payne; Adam Fourney; Amanda Swearngin; Wenyue Hua; Tori Westerhoff; Amanda Minnich; Maya Murad; Ece Kamar; Ram Shankar Siva Kumar; Saleema Amershi
Reports a red-team exercise on a live internal platform with more than 100 persistent agents and identifies propagation, amplification, trust capture, and invisibility failures.
Demonstrates interaction risks that do not appear in isolated-agent tests; the environment and findings are company-reported rather than independently reproduced.
Practitioner analysis
Reliability
age-0163
Research Foundations
Hypergraph topology optimization
Paper
HyperAgent: Leveraging Hypergraphs for Topology Optimization in Multi-Agent Communication
https://www.ifaamas.org/Proceedings/aamas2026/pdfs/QTVF9552.pdf
AAMAS
2,026
Heng Zhang; Yuling Shi; Xiaodong Gu; Zijian Zhang; Haochen You; Lubin Gan; Yilei Yuan; Jin Huang
Represents collaboration groups as hyperedges and uses a variational hypergraph autoencoder with sparsity regularization to generate task-adaptive communication topology.
Adds group-level topology learning beyond pairwise edges and differs from evolving-hypergraph work by focusing on generated sparse structures for each task.
Peer-reviewed research
Evolution
age-0164
Research Foundations
Three-layer knowledge sharing
Paper
D³MAS: Decompose, Deduce, and Distribute for Enhanced Knowledge Sharing in Multi-Agent Systems
https://www.ifaamas.org/Proceedings/aamas2026/pdfs/WQMU8577.pdf
AAMAS
2,026
Heng Zhang; Yuling Shi; Xiaodong Gu; Haochen You; Zijian Zhang; Lubin Gan; Yilei Yuan; Jin Huang
Connects task decomposition, collaborative reasoning, and distributed memory as coordinated layers in a heterogeneous graph designed to reduce duplicated retrieval, inference, and assignment.
Unifies work, reasoning, and state dependencies in one inspectable organization rather than optimizing those coordination surfaces separately.
Peer-reviewed research
Work graphs
age-0165
Protocols & Handoffs
Resource-bounded communication calculus
Paper
µACP: A Formal Calculus for Expressive, Resource-Constrained Agent Communication
https://www.ifaamas.org/Proceedings/aamas2026/pdfs/PHRW6922.pdf
AAMAS
2,026
Arnab Mallick; Indraveni Chebolu
Formalizes resource-bounded agent communication with PING, TELL, ASK, and OBSERVE operations, proves finite-state expressiveness and message bounds, and checks properties with TLA+ and Coq.
Provides a precise substrate for reasoning about protocol behavior and communication budgets; it is a general multi-agent calculus rather than an LLM-specific runtime.
Peer-reviewed research
Handoffs
age-0166
Research Foundations
Reinforced evolving orchestration
Paper
Multi-Agent Collaboration via Evolving Orchestration
https://papers.nips.cc/paper_files/paper/2025/hash/f1320d2e2842169c6fc89dcbd80e94d0-Abstract-Conference.html
NeurIPS
2,025
Yufan Dang; Chen Qian; Xueheng Luo; Jingru Fan; Zihao Xie; Ruijie Shi; Weize Chen; Cheng Yang; Xiaoyin Che; Ye Tian; Xuantang Xiong; Lei Han; Zhiyuan Liu; Maosong Sun
Trains a centralized orchestrator with reinforcement learning to select and sequence agents from evolving task state, producing dynamic inference graphs and compact cyclic reasoning paths.
Learns runtime activation and ordering rather than fixing a workflow in advance, providing a centralized contrast to decentralized evolutionary coordination.
Peer-reviewed research
Evolution