Six AWS announcements in three days describe the same architecture from six angles. Amazon Bedrock AgentCore is in general availability, and the pieces to move an agent beyond demo are all there: an MCP bridge connecting cloud agents to local tools on your laptop, persistent memory across sessions, an n8n node that builds agentic workflows without code, and native Web Search inside Bedrock for grounding without external providers.

For anyone trying to run an agent on real work, the perspective shift matters. The central question moves from what to how to orchestrate who does what. Which model to pick falls to the background. The agent becomes a process with state, tools, isolation, and observability. AWS brings these pieces together in a single platform, doing it the same week Cloudflare builds its runtime for agents. Two providers with different infrastructure, same design.

Published use cases are concrete. Mobileye deployed a technical support agent bridging on-prem systems with the cloud. LendingTree has three coordinated agents guiding customers on mortgages while respecting financial compliance. These are real productions, with numbers and real constraints behind them.

In detail

The agent you tested in chat, the one that answers well if you give it the right prompt, has a structural problem when you move it to production: it loses state between calls, doesn’t remember what it did, has nowhere safe to run code, and every tool you add is a hand-coded integration. AgentCore tries to solve these problems in one place.

Persistent memory. The agent saves state across sessions. If it analyzed a document yesterday and you ask for a follow-up today, it doesn’t start over. That’s the difference between an assistant that works and one that restarts every time.

MCP bridge for local tools. The Model Context Protocol, which we covered in the 2.0 update, is the standard for connecting agents to instruments. AWS’s bridge solves a concrete problem: the agent runs in the cloud, but your files and tools live on your laptop. The solution is a signed WebSocket tunnel through a browser extension, no ports open or VPN config.

n8n as visual frontend. The open source node for n8n turns AgentCore into a step in a visual workflow. Memory, real tools, code execution, and VPC isolation configure from the editor, no infrastructure code. For non-developers, it means building an agent by dragging blocks.

Native Web Search. Bedrock now has integrated server-side web search. Model grounding on current sources becomes a platform capability, no third-party onboarding or external API management.

What real cases say. Mobileye bridges on-prem systems with AWS for technical support while keeping enterprise governance. LendingTree uses three coordinated agents with LangGraph and MCP for mortgages, with integrated guardrails for financial compliance. The third case, automated insight extraction from the web, combines AgentCore Browser, OpenSearch, and Lambda to monitor RSS feeds and make extracted insights searchable.

Limitations. These are AWS blog posts: they describe the architecture at its best angles. Lock-in is real (once built on AgentCore, switching providers means rewriting orchestration). The MCP bridge is elegant but adds complexity (browser extension, Chrome native messaging). And convergence with Cloudflare confirms direction without validating details: two providers converging on the same design is a strong signal, but it doesn’t replace field experience.