June 2025 Summaries

After extensive experience with authentication systems and developer tools, the author observes challenges in the Model Context Protocol (MCP) ecosystem, particularly focusing on the gap between experimental implementations and production-ready infrastructure. While MCP has significant potential for enabling AI agents to interact with real-world systems, there are security and scalability issues that need addressing, especially when MCP servers interact with external APIs. The current state of MCP security involves basic client-server authentication, but embedding admin-level credentials in servers creates security risks and engineering challenges. A proposed solution involves user-specific authorization flows, where servers obtain tokens that inherit user permissions, as demonstrated in PR #475. Compliance with security standards is crucial to avoid interoperability issues, security vulnerabilities, and scalability problems. The text stresses the importance of production readiness, which includes observability, scalability, error handling, rate limiting, and audit trails. Arcade.dev aims to provide a secure and scalable MCP runtime, emphasizing the necessity of applying proven security principles to AI infrastructure for safely interacting with real-world systems.

AI's potential in transforming business operations is constrained by its current inability to access and act upon real enterprise data securely. Despite advancements in large language models (LLMs) and their ability to understand context and make decisions, the actual value of AI in enterprise settings is realized only when these models can integrate with existing business systems. The Agent Hierarchy of Needs framework outlines the necessary layers for this integration, beginning with LLMs and including prompt orchestration, retrieval systems, agent orchestration, tool calling, and agent authorization, culminating in agentic action. This complex infrastructure requires secure authentication and granular permissions to ensure AI agents can perform meaningful tasks without compromising security or governance. Arcade.dev addresses these challenges by providing the necessary tools for secure system integration and authorization, enabling AI to execute complex business processes and move beyond mere pilot projects to full-scale deployments.

Arcade.dev has been recognized as the "Overall Authentication Solution of the Year" in the 8th Annual AI Breakthrough Awards for addressing a significant challenge in AI development: enabling AI agents to authenticate and interact with real-world tools effectively. The company's solution provides secure OAuth flows, over 50 pre-built connectors to popular platforms like Gmail and Slack, and a streamlined setup process, allowing AI agents to perform tasks they previously couldn't, like scheduling meetings or sending emails. This development underscores a shift in the AI industry towards building robust infrastructure that allows AI systems to function beyond theoretical applications and become practical tools in production environments. The award highlights the importance of authentication as a crucial component in making AI agents truly operational, reducing the gap between AI's potential and its real-world applicability.

The discussion explores "Agentic Delegation," a method in multi-agent systems where tasks are delegated between agents based on internal criteria, using frameworks such as LangGraph, OpenAI's Agents SDK, and Google's Agent Development Kit (ADK). Each framework implements agentic delegation through tool calling, transferring control to sub-agents. LangGraph offers a comprehensive experience by allowing fine-grained control and the ability to construct raw graphs, while OpenAI's SDK provides explicit handoff modeling with two approaches: full delegation of flow or input generation by the calling agent. Google's ADK, though not as flexible in handling complex contexts, efficiently manages most cases with a tool that specifies the target agent and invocation context. The choice of framework depends on the required level of control over the context, with LangGraph favored for detailed control, whereas Google ADK caters to those willing to let the orchestration framework manage context control.

The text discusses the fundamental differences between Retrieval-Augmented Generation (RAG) systems and true agents, emphasizing that while RAG systems are essentially enhanced search engines with natural language interfaces, true agents possess the capability to perform actions autonomously. It highlights that RAG systems are limited to providing information and require significant architectural changes to evolve into agents, which can execute tasks such as scheduling meetings or filing reports. The transformation from RAG to agents involves complex requirements such as tool calling infrastructure, state management, and sophisticated error handling, distinguishing them through their ability to manage workflows and take actionable steps based on context. The text argues against conflating the two and stresses the need for proper engineering to create agents that not only assist users but also execute tasks on their behalf, ultimately expanding the possibilities and usefulness of AI systems.

The blog post explores the evolving landscape of agentic frameworks, focusing on how different systems approach agent orchestration and Human-in-the-Loop (HITL) processes. The author experiments with three frameworks—LangGraph, OpenAI's Agents SDK, and Google's Agent Development Kit (ADK)—to implement a supervisor-based agentic system, each with distinct approaches to HITL. While LangGraph is praised for its sophisticated documentation and ease of implementation using graph states and interrupts, the ADK offers simplicity through callbacks, albeit with some ergonomic challenges. OpenAI's SDK, although valuable for deepening Python knowledge, is criticized for its awkward control flow mechanisms. The post emphasizes that the best framework depends on the user's goals, recommending LangGraph for production-ready projects and ADK for beginners due to its clarity in illustrating architectural patterns. The full experiment is open-source and accessible for further exploration.