\n ## Diagram: Multi-Agent Communication Protocols ### Overview The image is a technical diagram illustrating and contrasting four different communication protocols for multi-agent systems, specifically focusing on task assignment and evaluation workflows. It is divided into four quadrants, each depicting a distinct method with labeled agents, directional arrows representing communication flow, and explanatory text. The diagram uses color-coded agents (pink for Generator, green/blue for Evaluators, black for Supervisor) and numbered steps to show sequences. ### Components/Axes The diagram is structured into four main sections, each with a title and a visual workflow: 1. **Top-Left Quadrant: Context-Poor Communication** * **Title:** "Context-Poor Communication" * **Agents:** Supervisor (black), Member Agent 1 (Role: Generator, pink), Member Agent 2 (Role: Evaluator, green). * **Flow:** A linear, back-and-forth sequence between Supervisor and each agent individually. * **Key Text Elements:** * Prompt box: "Prompt (General Task Assignment)" * Supervisor to Agent 1: `① "Your sub task is...You need...." (Only-text based, Length Task)` * Agent 1 thought bubble: "How Many? In what Format? ..." * Agent 1 to Supervisor: `② "The output of mine is:..."` (Labeled: "Bad Response due to Non-organized instruction") * Supervisor to Agent 2: `③ "Your sub task is...You need...." (Only-text based, Length Task)` * Agent 2 thought bubble: "Metrics? Evaluate what? ..." * Agent 2 to Supervisor: `④ "The output of mine is:..."` (Labeled: "Bad Response due to Non-organized instruction") * **Explanatory Text:** "Context-Poor Communication: ① Task Assignment: Assigning the next agent without providing well-organized information. ② Bad Responses: Producing poor responses due to forgetting key issues, such as the context of the question or intermediate format." 2. **Top-Right Quadrant: Structured Communication** * **Title:** "Structured Communication" * **Agents:** Supervisor (black), Member Agent 1 (pink), Member Agent 2 (green). * **Flow:** Supervisor communicates with each agent using a structured protocol. * **Key Text Elements:** * Prompt box: "Prompt (General Task Assignment)" * Supervisor to Agent 1: `① Well-organized communication protocol` with sub-points: * `a. (Message With Specified Subtask): "Your Subtask is to ...the format is..."` * `b. (Intermediate Output): "..."` * `c. (Background): "We are discussing ..."` * Agent 1 thought bubble: "To finish this I need ...." * Agent 1 to Supervisor: `② Accurate Response with Intermediate Output` * Supervisor to Agent 2: `③ Well-organized communication protocol` (same structure as ①). * Agent 2 thought bubble: "These are good as..." * Agent 2 to Supervisor: `④ Accurate Response with Intermediate Output` * **Explanatory Text:** "Structured Communication: ① Task assignment: Assigning specific subtask to the next agent with *message, intermediate output and background*. ② Accurate responses: including response and other intermediate output." 3. **Bottom-Left Quadrant: Traditional Evaluation based Refinement** * **Title:** "Traditional Evaluation based Refinement" * **Agents:** Supervisor (black), Agent 1 (Role: Generator, pink), Agent 2 (Role: Evaluator 1, green), Agent 3 (Role: Evaluator 2, blue). * **Flow:** A strict, predefined sequential order: Supervisor → Agent 1 → Agent 2 → Agent 3 → Supervisor. * **Key Text Elements:** * Supervisor to Agent 1: `① Task: "Generate text...."` * Agent 1 to Agent 2: `② "My Generated text is...."` * Agent 2 to Agent 3: `③ "My evaluated result is...."` * Agent 3 to Supervisor: `④ "My evaluated result is...."` * **Explanatory Text:** "Traditional Evaluation based Refinement: All agents follow a predefined order, such as Supervisor → Evaluator 1 → Evaluator 2, or similar styles, without summarization or coordination. This leads to biased outcomes (sensitivity to evaluator order) or incorrect results (overlooking partial evaluations)." 4. **Bottom-Right Quadrant: Hierarchy Refinement** * **Title:** "Hierarchy Refinement" * **Agents:** Supervisor (black), Evaluation Team Supervisor (grey), Member Agent 1 (Role: Generator, pink), Member Agent 2 (Role: Evaluator 1, green), Member Agent 3 (Role: Evaluator 2, blue). * **Flow:** Hierarchical. The main Supervisor delegates to an Evaluation Team Supervisor, who coordinates multiple evaluators in parallel or structured sequence. * **Key Text Elements:** * Supervisor to Agent 1: `① "Generate text...."` * Agent 1 to Supervisor: `② " My generated text is...."` * Supervisor to Evaluation Team Supervisor: `③ "Evaluate the text ...."` * Evaluation Team Supervisor to Agent 2: `④-a " Evaluate it from {placeholder= Criterion 1}"` * Agent 2 to Evaluation Team Supervisor: `⑤-a" The Score of {placeholder= Criterion 1} is ....."` * Evaluation Team Supervisor to Agent 3: `④-b " Evaluate it from {placeholder= Criterion 2}"` * Agent 3 to Evaluation Team Supervisor: `⑤-b" The Score of {placeholder= Criterion 2} is ....."` * Evaluation Team Supervisor internal note: `⑥ "The {placeholder= Criterion 1} is good, but the {placeholder= Criterion 2} is not good"` * **Explanatory Text:** "Hierarchy Refinement: An evaluator team operates hierarchically, providing summarized and coordinated feedback, balancing diverse inputs, and mitigating biases, which collectively enhance overall task quality." ### Detailed Analysis The diagram systematically compares communication inefficiencies and solutions in multi-agent systems. * **Context-Poor vs. Structured Communication:** The top half contrasts unstructured prompts (leading to confused agents and "Bad Responses") with structured prompts that include message, intermediate output, and background context (leading to "Accurate Response"). * **Traditional vs. Hierarchy Refinement:** The bottom half contrasts a linear, sequential evaluation chain (prone to bias and order sensitivity) with a hierarchical model where a team supervisor coordinates multiple evaluators, synthesizes feedback, and provides a summarized evaluation. * **Agent Roles:** Consistently color-coded: Pink = Generator, Green/Blue = Evaluator, Black = Primary Supervisor, Grey = Evaluation Team Supervisor (in Hierarchy Refinement). * **Communication Patterns:** Arrows are strictly directional, showing the flow of tasks and responses. Numbering (①, ②, etc.) clarifies the sequence within each protocol. ### Key Observations 1. **Problem-Solution Pairing:** Each quadrant on the left (Context-Poor, Traditional) presents a problem, while the corresponding quadrant on the right (Structured, Hierarchy) presents a proposed solution. 2. **Information Density:** The "Structured" and "Hierarchy" protocols explicitly include more metadata (background, intermediate outputs, criteria) in their communication compared to their counterparts. 3. **Coordination Complexity:** The "Hierarchy Refinement" model introduces an additional supervisory layer (Evaluation Team Supervisor) to manage complexity and bias, which is absent in the "Traditional" linear model. 4. **Visual Metaphors:** Thought bubbles are used effectively to show agent confusion in the poor protocol and clarity in the structured one. ### Interpretation This diagram serves as a conceptual framework for designing robust multi-agent AI systems. It argues that the quality of agent collaboration is fundamentally dependent on the structure of their communication protocols. * **The Core Argument:** Unstructured, context-poor communication leads to failure. Effective multi-agent systems require **structured message formats** that provide necessary context and **hierarchical coordination** to synthesize diverse evaluations and mitigate individual biases. * **Underlying Principles:** The diagram promotes principles of **explicit context sharing**, **intermediate output tracking**, and **coordinated evaluation** as antidotes to common failure modes like poor task understanding and biased results. * **Practical Implication:** For engineers designing such systems, the diagram suggests moving away from simple, sequential agent chains and towards architectures with well-defined communication schemas and supervisory coordination layers to achieve higher quality and more reliable outcomes. The "Hierarchy Refinement" model is presented as the most sophisticated approach, balancing diverse inputs through structured coordination.