May 6, 2026
Yesterday, Priyanka Vergadia discussed a common architecture for multi-agent systems, where a supervisor agent orchestrates specialized subagents, including a human-in-the-loop, a pattern confirmed by JP Morgan's "Ask David" AI. Separately, Maryam Miradi detailed 17 ways to implement Retrieval Augmented Generation (RAG), ranging from naive to self-improving methods. Dr. Milan Milanović also highlighted Brooks's Law, explaining why adding engineers to a late project often delays it further due to increased communication overhead, a concept mirroring the O(n²) scaling challenges in multi-agent coordination.
🔧 Spec & SDK
No new items were posted yesterday related to protocol spec changes, SDK releases, or breaking changes that would require a code update in existing MCP servers.
Why it matters: The absence of updates in this area suggests no immediate actions are required for MCP server builders to adapt to new specifications or SDK versions.
🏗️ Builder Patterns
Priyanka Vergadia yesterday highlighted the O(n²) scaling problem inherent in multi-agent coordination, noting that point-to-point communication grows quadratically with the number of agents. This means ten agents require 45 integration points, and 20 agents need 190. She outlined the current state of agentic systems, where manual handoffs between tools like GitHub Copilot, Claude Code, and Codex lead to signal loss and latency. Vergadia proposed a properly architected system requires shared runtime state, a framework-agnostic mesh, and an O(n) bus architecture, where each agent connects once to a central communication layer. Further supporting this, Vergadia also reported that JP Morgan revealed the architecture behind "Ask David," their investment research AI agent. This system utilizes a supervisor agent for routing and orchestration, specialized subagents for tasks like retrieval and analytics, an LLM-as-judge for reflection, and a human-in-the-loop to ensure accuracy. This pattern, she notes, is emerging as a standard for enterprise-scale agent deployments. Meanwhile, Dr. Milan Milanović yesterday reinforced Brooks's Law, explaining that adding engineers to a late project often makes it later due to increased communication overhead and onboarding time, a dynamic similar to the scaling issues in multi-agent systems.
Why it matters: MCP server builders should consider adopting bus architectures or supervisor-subagent models to manage scaling complexity in multi-agent systems, learning from enterprise implementations like JP Morgan's, and recognize the communication overhead implications for both human and AI team structures.
📄 Research
Maryam Miradi, PhD, yesterday presented 17 different approaches to Retrieval Augmented Generation (RAG), categorizing them into core paradigms, retrieval strategies, and self-improving methods. She described foundational concepts such as Naive RAG for basic retrieve-and-generate pipelines, Advanced RAG for pre- and post-retrieval optimization, and Modular RAG for composable retrieval modules. Retrieval strategies included Hybrid RAG combining dense and sparse retrieval, Fusion RAG fusing multiple queries, HyDE RAG using hypothetical document embeddings, Hierarchical RAG with multi-level chunking, and Adaptive RAG that adjusts based on query complexity. Miradi also covered self-improving RAG techniques like Self-RAG, where the model critiques its own outputs, Corrective RAG for detecting and fixing bad retrievals, and Speculative RAG which drafts answers before verification. These advancements build on earlier work, including the RAG review paper, with version one submitted on December 18, 2023, and version five on March 27, 2024.
Why it matters: MCP server builders can leverage these advanced RAG patterns to improve the accuracy, efficiency, and adaptability of their agentic systems, particularly for information retrieval and generation tasks.
So What?
MCP server builders should evaluate their multi-agent system architectures for O(n²) scaling issues and consider implementing bus architectures or supervisor-subagent patterns for better coordination. Analyze JP Morgan's "Ask David" architecture as a proven enterprise model for multi-agent system design, integrating specialized subagents and a human-in-the-loop for enhanced reliability. Additionally, explore advanced RAG techniques, such as Hybrid or Adaptive RAG, to optimize information retrieval and generation within your MCP agents.
