AgentMaster: A Multi-Agent Conversational Framework Using A2A and MCP Protocols for Multimodal Information Retrieval and Analysis

1 Introduction

• Few applications had employed A2A and MCP together within a single multi-agent framework, despite challenges in coordination and interaction with heterogeneous tools and data.The introduction also identifies domain-specific expertise and multimodal data as additional challenges.Source: §1 · pp. 1–2
• AgentMaster integrates A2A and MCP to coordinate specialized agents, decompose queries, allocate tasks, and dynamically route workflows.A unified conversational interface supports natural-language interaction without requiring prior technical expertise.Source: §1 · p. 2
• The pilot prototype self-implements A2A and MCP and supports information retrieval, image analysis, database querying, question answering, and content summarization.It is deployed locally and on AWS as Flask-based microservices.Source: §1 · p. 2
• The evaluation uses G-Eval, BERTScore, and related metrics to assess correctness, completeness, and semantic fidelity across diverse query types.The passage describes evaluation measures but does not report their values.Source: §1 · p. 2

2 The General System Framework

• The framework has four components: a unified conversational interface, a multi-agent center, multi-agent protocols, and a state management layer.The architecture is presented as the general MAS framework of AgentMaster.Source: §2 · p. 3

2.1 Unified Conversational Interface

• The interface accepts text, charts, images, and audio, and can generate text, images, and structured data tables..Source: §2.1 · p. 3

2.2 Multi-Agent Center

• The multi-agent center combines an orchestrator, domain agents, and general agents, with A2A for agent communication and MCP for external tools and resources.Domain agents may be LLM-based or non-LLM-based; each general agent is paired with a dedicated LLM.Source: §2.2 · p. 3

2.2.1 Orchestrator Agent

• The orchestrator delegates tasks by agent capability, decomposes complex requests into subtasks, manages protocol errors, and synthesizes agent outputs.Subtasks may run in parallel or sequentially.Source: §2.2.1 · p. 3

2.2.2 General Agents

• The orchestrator selects a general agent for broad tasks without domain-specific dataset access, or a domain agent when specialized capabilities are appropriate.The choice depends on task nature and complexity.Source: §2.2.2 · p. 3

2.2.3 Domain Agents

• Domain agents connect to domain-relevant functions, datasets, and tools; AgentMaster describes SQL, information retrieval, and multimodal analytics as supported specializations.Domain agents can manage sub-agents and collaborate with one another and general agents.Source: §2.2.3 · p. 3

2.3 Multi-Agent AI Protocols

• A2A standardizes JSON-based agent communication, while MCP provides unified tool access, long-term memory, and context management.The framework may use Google's A2A SDK or implement A2A itself; MCP is developed similarly.Source: §2.3 · pp. 4–5

2.4 State Management Layer

• Vector databases provide persistent semantic memory, while context caches hold session data and intermediate results during active workflows.Together they maintain MCP state for multistep, user-specific, and domain-specific tasks.Source: §2.4 · p. 5

3 System Architecture of the Case Study

• The case-study architecture combines modular components for retrieval-augmented question answering through dynamic agent orchestration.The system architecture is illustrated in Figure 2.Source: §3 · p. 5

3.1 Coordinator Agent and Complexity Assessment

• The Coordinator assesses query complexity, routes simple requests to an MCP client, and sends complex requests through workflows involving multiple retrieval agents.It performs query analysis and orchestration.Source: §3.1 · p. 5

3.2 Agent Clients and MCP Clients

• Agent clients invoke distributed workflows, while MCP clients send JSON-RPC requests to MCP servers containing domain-specific retrieval logic.The division supports compositional retrieval and fallback handling without manual routing configuration.Source: §3.2 · p. 5

3.3 Retrieval Agents

• Four retrieval agents handle unstructured knowledge-base content, relational-database SQL, image inputs through vision APIs, and open-domain or fallback queries.Each agent exposes an MCP server endpoint.Source: §3.3 · p. 5

3.4 LLM Integration and Error Handling

• A local or external LLM synthesizes retrieval agents’ partial outputs, while the Flask server and Coordinator include error handling for retrieval and inference failures.The LLM also performs language generation, reasoning, and summarization.Source: §3.4 · p. 5

3.5 End-to-End Workflow

• Queries pass from the interface to the Flask server and Coordinator, then through retrieval agents and MCP clients before the LLM synthesizes the final response.The input may be text or an image.Source: §3.5 · p. 5

3.6 Design Considerations

• New retrieval agents can be added without changing orchestration logic, and standardized JSON-RPC interfaces support consistent agent communication.The authors describe the design as a flexible foundation for retrieval-augmented conversational systems.Source: §3.6 · p. 5

4.1 Individual Agent Evaluation

• Specialist agents were previously reported as reliable, but occasional misclassification of simple queries as complex produced incorrect responses.Individual single-task queries otherwise consistently yielded correct results in testing.Source: §4.1 · p. 6

4.2 Complex Task Evaluation

• Across six complex queries, decomposition was correct and most agent paths were assigned correctly; simple-query checks matched composite answers in the bridge-count example.The simple-query checks covered total bridges in Virginia and bridges built there in 2019.Source: §4.2 · p. 6, Table 1 · p. 6

4.3 Overall Evaluation

• Table 3 presents feature-level comparisons of AgentMaster with systems using neither protocol, A2A only, or MCP only.The passage characterizes the comparison as showcasing AgentMaster’s features with A2A and MCP.Source: Table 3 · p. 7, §4.3 · p. 7
• Across 23 simple and complex questions, complex-query G-Eval exceeds 87.1%, while BERTScore F1 averages 96.3%, reflecting high semantic fidelity to reference outputs.Table 2 reports aggregated metrics across all query types.Source: §4.3 · p. 7
• SQL and IR agents produced consistently accurate results, while the general and image agents showed minor variability; most complex-query sub-questions matched the composite responses.Source: §4.3 · p. 7

• AgentMaster uses A2A-MCP protocols for retrieval-augmented question answering across structured, unstructured, and multimodal data sources.Source: §5 · p. 7
• BERTScore F1 averages 96.3% and G-Eval averages 87.1%, with human and quantitative evaluation supporting coordination and semantically faithful responses.Source: §5 · p. 7
• The authors present agent-based orchestration as having potential for scalable, domain-adaptive conversational AI.Source: §5 · p. 7

• Retrieval and generation accuracy are partly constrained by the underlying LLM and retrieval corpus.Source: §6 · p. 7
• Occasional query-complexity misclassification can cause unnecessary decomposition or incomplete responses, while limited collaboration and database size can restrict informational depth.LLM reasoning may also have difficulty synthesizing complex information.Source: §6 · p. 7
• LLM-as-a-judge may be biased, lack task-specific expertise, or misalign with human judgment, and the framework lacks established safeguards for information storage and usage.Source: §6 · p. 7

• Appendix demonstrations include multimodal inputs and outputs.Source: A Appendix · p. 10
• The demonstrations cover single and complex queries, including verification and uses of domain-specific or specialized agents.Source: A Appendix · p. 10
• Backend examples show query decomposition, task allocation, and dynamic routing for complex queries.Source: A Appendix · p. 10

• Figures 4–5 show frontend examples of single queries for the Image and IR agents, respectively, while Figures 6–7 show additional single-query demonstration examples for the SQL agent.Source: Figure 4 · p. 10, Figure 5 · p. 10, Figure 6 · p. 11
• Figures 8–9 show two examples each of verification queries for the IR and image agents.Source: Figure 8 · p. 13, Figure 9 · p. 14
• Figure 10 highlights duplicate content among errors in SQL information retrieval, indicating AgentMaster may still produce erroneous results.Source: Figure 10 · p. 15
• Figure 11 provides an additional AgentMaster demonstration and verification example.Source: Figure 11 · p. 16

A.2.1 Complex Queries

• Table 4 lists the complex questions and their corresponding IDs, matching Table 1.Source: Table 4 · p. 17

A.2.2 AgentMaster Outputs of Complex Queries in the Chat Interface

• Figures 12–14 display frontend responses for complex queries Q1–Q6.Source: Figure 12 · p. 17, Figure 13 · p. 18, Figure 14 · p. 19

A.2.3 Automated Actions (Query Decomposition, Task Allocation, and Dynamic Routing) for Complex Queries in the Backend

• Tables 5 and 6 document backend automated actions for complex queries Q1–Q3 and Q4–Q6, respectively.Source: Table 5 · p. 20, Table 6 · p. 21