Paolo Perrone Redraws AI Agent Stack Into 6 Layers for 2026 as MCP and Eval Rise
Updated
Updated · O'Reilly Media · Jun 8
Paolo Perrone Redraws AI Agent Stack Into 6 Layers for 2026 as MCP and Eval Rise
1 articles · Updated · O'Reilly Media · Jun 8
Summary
A new 2026 diagram recasts the AI agent stack into six layers, arguing teams should start with the simplest workable setup and add complexity only when a specific production problem appears.
Three shifts drove the redraw from the 2024 map: MCP created a distinct tools layer, reasoning models replaced some multistep chains with single-call agents, and memory became a first-class architectural layer.
The framework update also elevates eval and guardrails as major production bottlenecks, citing a LangChain survey showing 89% of teams have observability but only 52% have evals.
At the tools layer, Perrone says MCP has effectively won standardization—with 97 million monthly SDK downloads and backing from OpenAI, Google and Microsoft—but security remains the main unresolved risk.
The broader argument is that many teams still build like it is 2024, overusing LangGraph, vector databases and multi-agent designs before they need them, even as provider SDKs may absorb much of the stack by early 2027.
As provider SDKs consume the agent stack, will custom builds become a competitive advantage or a technical debt trap?
With agent security being the least mature layer, what major exploit are we overlooking that could trigger a widespread crisis?
If an agent's identity now lies in its memory, who will ultimately own and control this valuable personalized data?
The 2026 AI Agent Stack: Architecture, Security, Evaluation, and the Race for Safe, Scalable Autonomy
Overview
In response to the fragmented and chaotic development of AI agents seen in 2024 and 2025, Paolo Perrone introduced the '2026 Redraw' on June 5, 2026. This new six-layer architecture was created to address the shortcomings of earlier, ad-hoc agent designs and the compatibility issues caused by an explosion of disparate frameworks. The 2026 Redraw provides a unified and scalable foundation for AI agents, aiming to overcome past limitations and growing pains by emphasizing modularity, interoperability, and a structured approach to building reliable and adaptable systems.